Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

DOI： 10.1016/j.enconman.2026.121790

Authorship：Last author,　Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

Fault diagnosis in heating, ventilation, and air-conditioning (HVAC) systems is essential for intelligent energy management. However, extreme class imbalance often degrades diagnostic performance. Existing data augmentation methods typically assume a fixed and known set of fault categories; when new fault categories emerge, these methods may become ineffective. To address this challenge, we propose an adaptive augmentation framework (ADAF) based on a meta-diffusion model. Specifically, we develop a meta-diffusion denoising implicit model (MDDIM) that is meta-trained on known fault categories and can rapidly adapt to imbalanced datasets containing new fault categories, enabling the generation of realistic fault samples. To better capture the distribution of an emerging fault category from scarce data, we further design a prototype-similarity-based strategy to initialize the class condition vector and integrate it with a conditional noise prediction network (CNPN) to improve generation quality. Extensive experiments on three HVAC datasets demonstrate that the proposed method consistently outperforms seven representative augmentation baselines. Across different imbalance settings, MDDIM improves the F1 score by 2.75%–21.75% after augmentation, providing an effective solution for fault diagnosis under extreme imbalance with emerging fault categories.

DOI： 10.1016/j.apenergy.2026.128192

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Advanced Engineering Informatics  

Accurate forecasting of solar irradiation is crucial for efficient system utilization, especially given the large number of newly established solar PV systems. Data-driven models, particularly deep learning models such as Long Short-Term Memory (LSTM) networks, have proven effective for solar PV power generation forecasting. However, limited training data in newly established systems often hinders the deployment of these models. To address this challenge, this study adapts an LLM-based reprogramming framework, named Multivariate TimeLLM, to few-shot, day-ahead solar-irradiation forecasting under data-scarce conditions representative of newly deployed PV systems. The pre-trained large language model (LLM) backbone is used without parameter tuning, and prior knowledge embedded in its weights is transferred to the forecasting task to improve data efficiency and enable faster deployment with less data. The proposed framework is evaluated using actual measured data from Tokyo under seasonal data-scarcity scenarios. It outperformed conventional LSTM and Transformer-GRU baselines, achieving mean squared error (MSE) reductions of 27.7% and 21.4%, respectively, when trained on limited seasonal data. These results highlight the predictive accuracy and practical relevance of the proposed framework for newly established PV systems under data-constrained conditions.

DOI： 10.1016/j.aei.2026.104765

Scopus

Authorship：Last author,　Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

Fault diagnosis of heating, ventilation, and air conditioning (HVAC) systems is pivotal for energy conservation and emissions reduction in smart-city energy infrastructure. Although existing HVAC fault diagnosis methods have advanced considerably, most of them require large amounts of fault data and assume identical label spaces between the training and test sets. With limited fault samples and the presence of unknown faults, conventional frameworks often experience marked performance degradation. To address these challenges, a few-shot unknown fault diagnosis framework with entropy-based uncertainty guidance (FEUG) is proposed for HVAC systems. First, to enhance the representational capacity and reusability of features, a feature association fusion module (FAFM) is developed. Building on FAFM, a multi-scale feature interaction association network (MFIAN) is designed to achieve correlation fusion across features of different scales, thereby strengthening joint attention to global and local information. Second, an adaptive contrastive learning module, an entropy-assisted classification module, and a prototype similarity module are proposed to improve intra-class compactness and inter-class separability. Lastly, by integrating entropy-based uncertainty with the few-shot learning mechanism, the FEUG significantly enhances performance in HVAC few-shot unknown fault diagnosis tasks. Comprehensive experiments and ablation analyses conducted on three HVAC datasets and against seven comparative methods demonstrate that FEUG consistently achieves superior accuracy across various scenarios (SC). Specifically, FEUG attains average F1 scores of 69.92%, 82.39%, and 78.48% on SZVAV (SC2), SZCAV (SC8), and chiller (SC12), respectively, exceeding the second-best models by 3.24, 5.39, and 2.35 percentage points.

DOI： 10.1016/j.aei.2026.104730

Scopus

Authorship：Last author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

Air handling units (AHUs) are the primary energy-consuming components within heating, ventilation, and air conditioning (HVAC) systems. Therefore, fault diagnosis of AHUs is crucial for ensuring indoor comfort and improving energy efficiency. Although deep learning-based methods have achieved great success in fault diagnosis of AHUs, existing methods require a large amount of labeled operational data for model training. However, acquiring high-quality labeled data demands substantial domain expertise and prohibitive annotation costs, leaving a vast majority of collected operational data unlabeled and hindering the practical deployment of deep learning-based methods. To overcome this bottleneck, this paper proposes a knowledge and data fusion-driven self-supervised learning (KDFD-SSL) method for the fault diagnosis of AHUs with limited labeled data. Specifically, the top- k important physical fault features are selected according to the prior diagnostic knowledge of AHUs. Meanwhile, a stacked autoencoder (SAE) is designed to extract high-level fault features from raw operational data, capturing complementary fault information to the prior diagnostic knowledge. Furthermore, a Transformer-based encoder is built within a self-supervised framework and pre-trained via knowledge-data fusion, which enables the extraction of discriminative and targeted features from unlabeled data, thereby reducing the reliance on labeled data for downstream diagnosis tasks. Experiments on two AHU datasets demonstrate the effectiveness and superiority of KDFD-SSL over other state-of-the-art fault diagnosis methods, particularly with limited labeled data. The proposed KDFD-SSL method provides a viable pathway for transitioning deep learning-based fault diagnosis from data-rich theoretical scenarios to data-scarce, real-world building energy management applications.

DOI： 10.1016/j.enbuild.2026.117646

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Missing data poses a critical challenge in the modeling and control of HVAC systems, where reliable time-series information is essential for energy optimization and fault detection. While deep learning has advanced imputation accuracy, existing models often struggle with robustness under high missing rates or require extensive fine-tuning. This study introduces ImputeLLM, a novel imputation framework that integrates a frozen large language model (LLM) encoder with a Transformer-based decoder and an adaptive hybrid loss function. Without any fine-tuning or prompt engineering, the model efficiently encodes masked time-series data into semantic embeddings and reconstructs missing values with high accuracy. The framework is validated using real-world monitoring data from a central cooling plant in Qingdao, China, under MCAR, MAR and MNAR masking conditions. Compared to conventional methods, the proposed approach achieves up to 37.7% improvement in MAPE over linear interpolation and shows 16.5% gain over traditional MAE-based losses. Furthermore, it demonstrates strong generalization across varying missing rates and feature observability levels. These results highlight the potential of LLM-based architectures for practical deployment in energy systems with noisy or incomplete data.

DOI： 10.1016/j.apenergy.2026.127822

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

The integration of Building-Integrated Photovoltaic and Battery (BIPVB) systems is a critical pathway toward carbon-neutral buildings, yet their optimal control is challenged by the stochastic nature of renewable generation and complex load dynamics. While Deep Reinforcement Learning (DRL) has emerged as a promising model-free control approach, traditional value-based and policy-gradient methods often suffer from training instability and lack explicit long-horizon optimization capabilities. This study explores the application of the Decision Transformer (DT) to Building-Integrated Photovoltaic and Battery (BIPVB) control by reformulating the task as a return-conditioned sequence modeling problem. By leveraging the Transformer architecture, the proposed method treats the control problem as a supervised learning task, predicting optimal actions based on historical states and desired future returns (Return-to-Go). We conduct a comprehensive benchmark using the CityLearn dataset, comparing DT against a Rule-Based Controller (RBC) and six state-of-the-art DRL algorithms, including SAC, PPO, and TD3. Simulation-based benchmark results show that the DT agent achieves the lowest operational cost across the tested scenarios, yielding an 18.5% cost reduction compared to RBC and outperforming the strongest DRL baseline. Furthermore, the DT shows competitive and relatively consistent performance under varying initial battery conditions. These results suggest that sequence-modeling approaches are a promising direction for offline building energy control and merit further investigation in broader settings.

DOI： 10.1016/j.energy.2026.141616

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Accurate ultra-short-term photovoltaic (PV) power forecasting is critical for the secure and efficient operation of power systems with high levels of renewable generation. Rapid cloud-cover variability, however, induces large and sudden fluctuations in PV output, which existing models struggle to capture and for which they provide limited measures of uncertainty. In this work, we propose PV-MM-Diffusion, an end-to-end multimodal diffusion model that jointly generates multiple future sky images and corresponding PV power distributions via two coupled denoising autoencoders. A cross-modal attention mechanism fuses complementary information from sky imagery and historical PV time series, and an empty-frame placeholder strategy allows the model to operate with image-only, PV-only, or combined inputs. Experiments on a 30 kW rooftop system show that PV-MM-Diffusion substantially improves probabilistic forecasting: it achieves a continuous ranked probability score (CRPS) of 2.63 kW (an 18 % reduction relative to the SkyGPT→U-Net baseline) and a Winkler score (WS) of 21.46 kW (a 40 % reduction). The model delivers tighter and more reliable prediction intervals, especially during extreme ramp events and rapid cloud transitions. These results demonstrate the promise of diffusion-based multimodal frameworks for flexible, uncertainty-aware PV integration in future low-carbon power systems.

DOI： 10.1016/j.apenergy.2026.127816

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Heating, ventilation, and air conditioning (HVAC) systems often operate with scarce fault labels and limited computational resources, posing challenges for reliable fault detection and diagnosis (FDD). Existing FDD studies largely rely on fully supervised data or post-hoc alarm aggregation, treat FDD as static classification without considering temporal dependencies, and employ complex backbones without evaluating deployment efficiency. Moreover, common contrastive learning (CL) augmentations such as scaling or permutation violate HVAC physical constraints, erasing magnitude anomalies critical for diagnosis. To address these limitations, this study reframes HVAC FDD as a multivariate time-series representation learning problem and proposes a contrastive self-supervised framework coupling a lightweight temporal encoder with a compact classifier. A physics-consistent strategy—combining timestamp masking and partially overlapping cropping—constructs positive pairs without destroying magnitude or channel semantics, while a hierarchical dual contrastive loss aligns same-timestamp embeddings and separates cross-sequence states across multiple resolutions. The resulting encoder–SVM architecture explicitly targets deployability, achieving high diagnostic accuracy with up to 90–97% less memory and 20–25% faster training than Transformer baselines. Experiments on the MZVAV AHU dataset with rigorous day-level splits show consistent superiority over recurrent, linear, and Transformer-based models, improving diagnostic accuracy by 20–30% and macro-F1 by 40–50%. This work delivers a label-efficient, physics-consistent, and deployment-ready framework for automated FDD in real-time building management systems.

DOI： 10.1016/j.apenergy.2026.127557

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Engineering Applications of Artificial Intelligence  

With the large-scale integration of photovoltaic (PV) resources introducing uncertainty and randomness to both the user and grid sides, accurately quantifying this uncertainty is critical for maintaining power grid stability and optimizing energy system operation. While scenario generation methods have advanced in addressing these challenges, they frequently rely on extensive historical data for training—an often scarce resource during the planning phase. To alleviate this limitation, this study proposes a K-medoids-enhanced Wasserstein Generative Adversarial Network (K-WGAN) integrated with feature clustering, which improves scenario generation accuracy and effectiveness and exhibits robust performance even with limited historical datasets. In comparative analyses with Latin hypercube sampling (LHS), K-WGAN (equipped with feature clustering) showed significant superiority: (1) It captured PV output characteristics more precisely, with generated scenario mean increasing by 5 % and variance rising by 13 % (relative to reference values); (2) As the training dataset size decreases from 100 % to 4 %, LHS outperforms WGAN overall under data scarcity with random sampling, especially in terms of mean and standard deviation. However, experiments using LHS for data sampling across ten groups demonstrate that WGAN exhibits superior performance, compared with random sampling; (3) Ablation experiments validated the contributions of K-medoids clustering, Wasserstein distance, and GAN structure, with the integrated model reaching 99 % prediction interval coverage probability (PICP). (4) Cross-regional validation using Japan and China datasets confirmed its adaptability to diverse climates and PV systems, yielding mean deviation '7 % and coverage rate '98 %. These results illustrate K-WGAN supports energy system planning under data scarcity while balancing prediction accuracy and computational efficiency.

DOI： 10.1016/j.engappai.2025.113647

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Renewable Energy  

This research introduces a reinforcement learning optimization framework for renewable energy systems, aimed at advancing Net-Zero Energy Buildings integrated with solar photovoltaic, biomass power generation, and battery storage. To address the challenges posed by mixed action spaces in the deployment of reinforcement learning, an algorithm utilizing a parameterized action space has been employed. This study is capable of managing the operational scheduling of various renewable energy sources without incurring additional computational load, thereby achieving Net-Zero Energy Buildings. The proposed model has been case-analyzed based on actual measurement data from existing energy systems. The study's findings indicate that the reinforcement learning algorithm with a parameterized action space, compared to the baseline model, can enhance off-grid operational performance by 4 %, offering a more promising route towards achieving Net-Zero Energy Buildings. Simultaneously, the time the battery operates within the safe range has increased by 90 % compared to the baseline model, enhancing the system's energy flexibility. While achieving these objectives, there has been no additional computational burden on the reinforcement learning algorithm. This provides a feasible approach for the zero-carbon operation of office buildings and offers guidance and reference for stakeholders looking to develop similar carbon-neutral structures.

DOI： 10.1016/j.renene.2025.124493

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

Fault diagnosis in heating, ventilation, and air conditioning (HVAC) systems is crucial for maintaining energy efficiency and reducing carbon emissions in data centers. Most existing data-driven HVAC fault diagnosis approaches depend on a sufficient quantity of labeled or unlabeled operational data. However, newly constructed data centers often lack both labeled and unlabeled HVAC operational data. Moreover, a review of the literature reveals that few-shot learning has received limited attention in the context of HVAC fault diagnosis for data centers with an extremely limited quantity of operational data. In this study, a semi-supervised adaptive weighted prototype network (SSAWPN) incorporating an attentive feature fusion approach is proposed for HVAC fault diagnosis in data centers with minimal data. First, a multi-scale attentive feature fusion network (MSAFN) leverages channel-wise segmentation, residual connections, and an attention mechanism to capture fault signatures across multiple spatial and temporal scales. Then, a semi-supervised adaptive weighted prototype optimization strategy (SAWPS) is employed to incrementally update class prototypes by assigning adaptive weights to unlabeled data. As new data accumulate, the prototypes become increasingly representative of actual fault modes without manual annotation. Lastly, real-world operational data from the chiller and air handling unit (AHU) are used to conduct ablation and comparative experiments. The experimental results show a clear advantage for SSAWPN in HVAC few-shot settings. It achieves mean F1 scores of 73.77 % on the ASHRAE RP1043 chiller fault severity level 1 dataset and 67.22 % on the ASHRAE RP1312 AHU summer dataset, outperforming the second-best approach by 3.21 and 1.35 percentage points, respectively.

DOI： 10.1016/j.apenergy.2025.127056

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

High-quality photovoltaic (PV) power forecasting is essential for efficient energy management and reliable grid integration, yet real-world data are often plagued by extensive missingness in both target and auxiliary variables. To address this challenge, we propose MDCTL-MCI, a missingness-aware forecasting framework that jointly leverages signal decomposition, multi-scale covariate interaction, and multi-domain collaborative transfer learning. First, multivariate singular spectrum analysis (MSSA) denoises and reconstructs incomplete time series, enhancing underlying temporal structures without explicit imputation. Next, a lightweight multiscale covariate interaction (MCI) module models interactions among reconstructed PV power, global horizontal irradiance, direct normal irradiance, and total solar irradiance at varying temporal resolutions, capturing both local fluctuations and global trends. Finally, a multi-source domain collaborative transfer learning strategy aggregates knowledge from multiple PV sites to form a global model, which is then fine-tuned on a small set of high-quality, MSSA-processed samples at each site. By freezing all but the output layer during fine-tuning, MDCTL-MCI adapts efficiently to local data heterogeneity. Extensive experiments on four Chinese PV installations reveal that, compared to baseline methods, the proposed method improves average accuracy by 10.5 % under complete data conditions and by 15.3 % under various missing data scenarios.

DOI： 10.1016/j.apenergy.2025.126771

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

Data-driven fault detection and diagnosis (FDD) methods are emerging and attractive techniques for smart energy management in buildings, including the energy management in heating, ventilation, and air conditioning (HVAC) sub-systems. However, the real-world deployment of FDD in HVAC is hindered by data unavailability scenarios. In the past few years, various data augmentation methods, such as the generative adversarial network (GAN), have been proposed to address the abovementioned problem. However, these data augmentation methods suffer from stability, reliability, and interpretability issues. This paper proposes an interpretable ensemble learning-based diffusion model (IELDM) for HVAC systems, generating stable, reliable synthetic datasets to address the data unavailability issue. A split-gain-based method is introduced in IELDM to enhance the interpretability of the overall machine learning framework. Experimental results show that IELDM stably boosts FDD accuracy under extremely limited fault data, with improvements of up to 11.2 %, 13.2 %, and 12.08 % across three HVAC systems, clearly outperforming current state-of-the-art methods. By systematically overcoming the challenges of instability, unreliability, and lack of interpretability in current generative models, this work offers a robust solution to close the application gap of HVAC FDD in practical building energy systems.

DOI： 10.1016/j.apenergy.2025.126774

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Sustainability Switzerland  

Buildings with electrified heat pump systems, onsite photovoltaic (PV) generation, and energy storage offer strong potential for demand flexibility. This study compares two storage configurations, thermal energy storage (TES) and battery energy storage (BESS), to evaluate their impact on cooling performance and cost savings. A Model Predictive Control (MPC) framework was developed to optimize system operations, aiming to minimize costs while maintaining occupant comfort. Results show that both configurations achieve substantial savings relative to a baseline. The TES system reduces daily operating costs by about 50%, while the BESS nearly eliminates them (over 90% reduction) and cuts grid electricity use by more than 65%. The BESS achieves superior performance because it can serve both the controllable heating, ventilation, and air conditioning (HVAC) system and the home’s broader electrical loads, thereby maximizing PV self-consumption. In contrast, the TES primarily influences the thermal load. These findings highlight that the choice between thermal and electrical storage greatly affects system outcomes. While the BESS provides a more comprehensive solution for whole-home energy management by addressing all electrical demands, further techno-economic evaluation is needed to assess the long-term feasibility and trade-offs of each configuration.

DOI： 10.3390/su17219497

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Renewable Energy  

Accurate forecasting of electricity demand and generation is crucial for efficient grid management and sustainable energy planning. While large language models (LLM) have shown promise in various fields, their application to time series forecasting presents challenges, including limited cross-channel information capture and the complexity of prompt design. In this study, we propose a novel framework that combines multiple attention mechanisms with LLM, enabling effective feature extraction from both target and non-target variables without the need for prompt engineering. We conducted extensive experiments using real-world electricity demand and generation data from multiple regions in Japan to evaluate the proposed model. The results demonstrate that our model outperforms state-of-the-art LLM-based and other time-series forecasting models in terms of electricity demand and generation forecast task, achieving better performance than the latest LLM-based models without using prompts or increasing model size. Compared with the Long short-term memory network (LSTM), the mean absolute error (MAE) is reduced by 20.8%. Compared with the previous time-series LLM, the proposed model reduces memory usage by 49.3% and shortens training time by 35.7%. Additionally, the proposed model exhibits superior generalization ability, maintaining high performance even in zero-shot learning scenarios. Compared with the LSTM, MAE on the four test datasets is reduced by 16.6%.

DOI： 10.1016/j.renene.2025.123465

Scopus

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Energy and Buildings  

Residential buildings account for a significant portion of global energy consumption, making the optimal control of air conditioning and energy systems crucial for improving energy efficiency. However, existing reinforcement learning (RL) methods face challenges, such as the need for carefully designed reward functions in direct RL and the dual training phases required in imitation learning (IL). To address these issues, this study proposes a Generative Adversarial Soft Actor-Critic (GASAC) framework for controlling residential air conditioning and photovoltaic-battery energy storage systems. This framework eliminates the need for predefined reward functions and achieves optimal control through a single training process. An accurate simulation model was developed and validated using real-world data from a residential building in Japan to evaluate the proposed method's performance. The results show that the proposed method, without requiring a reward function, increased the time the temperature remained within the comfort range by 11.43 % and reduced electricity costs by 14.05 % compared to baseline methods. Additionally, the training time was reduced by approximately two-thirds compared to direct RL methods. These findings demonstrate the effectiveness of GASAC in achieving optimal temperature control and energy savings while addressing the limitations of traditional RL and IL methods.

DOI： 10.1016/j.enbuild.2025.115564

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

The deployment of renewable energy and the implementation of intelligent energy management strategies are crucial for decarbonizing Building Energy Systems (BES). Although data-driven Deep Reinforcement Learning (DRL) has achieved recent advancements in optimizing BES, significant challenges remain, such as the lack of studies addressing the observation space of time series data and the scarcity of interpretability. This paper first introduces future forecast information into the DRL algorithm to form the observation space for time series data. It employs Gated Recurrent Unit(GRU) and Transformer networks coupled with the DRL algorithm for operational control of a Building-Integrated Photovoltaic and Battery(BIPVB) system. Additionally, it aims to enhance the interpretability of the model regarding global and local feature importance by integrating the state-of-the-art Shapley Additive Explanations (SHAP) technique with the developed DRL model. All results were validated and tested on an open-source, real-world BIPVB system, showing that incorporating forecast information can reduce operational costs by 3.56%, while using GRU and Transformer networks to handle time-series data can further reduce costs by over 10%. The results of the SHAP value analysis demonstrated the importance of future electricity prices in forecast information for optimization, revealing the model's complex nonlinear relationships. Additionally, this study provided interpretability for a single episode instance based on the SHAP method. Overall, the study offers an accurate, reliable, and transparent deep reinforcement learning model, along with an insightful framework for handling time-series observations in DRL.

DOI： 10.1016/j.apenergy.2025.125387

Scopus

Authorship：Lead author   Publishing type：Research paper (scientific journal)   Publisher：Renewable Energy  

’Demand-side response’ (DSR), a mechanism through which residential electricity usage adapts based on external cues, has been conceptualized diversely, with numerous experiments showing that DSR frequently results in disparate and inconsistent outcomes for consumers. This diversity in outcomes prompts an examination of implementing such policies, thereby situating them within the discourse of energy justice—a perspective that explores the ethical dimensions of energy systems. However, current research lacks quantitative assessments and case studies specifically focused on energy justice. This article utilizes actual residential energy consumption data from Texas, USA, and designs scenarios based on a hypothetical renewable energy system and real-world DSR conditions. It quantitatively models the operational costs and peak loads resulting from different groups’ energy behaviors within the DSR framework. The results indicate that users with light participation in DSR can achieve up to 50% savings in operational costs on certain typical days, and as the level of DSR participation increases, these savings can reach approximately 90%. Users who are unable to participate in DSR are often those most vulnerable to energy poverty. Such policies thus pose a significant risk of energy injustice. Finally, based on the quantitative analysis of this energy injustice, we provide corresponding policy recommendations.

DOI： 10.1016/j.renene.2025.122477

Scopus

Authorship：Lead author,　Corresponding author   Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Deep learning models are increasingly being used to predict renewable energy-related variables, such as solar radiation and outdoor temperature. However, the black-box nature of these models results in a lack of interpretability in their predictions, and the design of deep network architectures significantly impacts the final prediction outcomes. The introduction of Kolmogorov–Arnold Network (KAN) provides an excellent solution to both of these issues. We hope that the KAN mechanism can provide fully interpretable neural network models, enhancing the potential for practical deployment. At the same time, KAN is capable of achieving good prediction results across various network architectures and neuron counts. We conducted case studies using real-world data from the Tokyo Meteorological Observatory to predict solar radiation and outdoor temperature, comparing the results with those of commonly used recurrent neural network baseline models. The results indicate that KAN can maintain model performance regardless of the chosen number of neurons. For instance, in the solar radiation prediction task, the KAN with a single hidden neuron reduces the MSE error by 75.33% compared to the baseline model. More importantly, KAN allows for the quantification of each step in the network's computations, thereby enhancing overall interpretability.

DOI： 10.1016/j.apenergy.2024.124844

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

The rising demands for comfort alongside energy conservation underscore the importance of intelligent air conditioning control systems. Model Predictive Control (MPC) stands out as an advanced control strategy capable of addressing these demands. However, accurate prediction of all relevant variables remains a challenge in practical scenarios, complicating MPC's ability to devise effective control actions amid prediction inaccuracies. To counteract this issue, this paper introduces an enhanced Double-Layer Model Predictive Control (DLMPC) algorithm. This innovative approach adjusts for discrepancies between forecasted and actual values without the need for additional variables and models, thereby reducing the adverse effects of prediction errors. Additionally, we develop precise models for room temperature simulation and for calculating air conditioning (AC) load and energy consumption, grounded in empirical data from residential settings and AC performance tests. Validation of these models demonstrates their efficacy in enabling MPC to formulate efficacious control strategies. When juxtaposed with a baseline model, the DLMPC algorithm significantly improves temperature regulation accuracy by up to 15.12% and achieves a 10.50% reduction in energy consumption over the heating season.

DOI： 10.1016/j.apenergy.2024.123652

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

The use of Deep Reinforcement Learning (DRL) in building energy management is often hampered by data efficiency and computational challenges. The long training time, unstable, and potentially harmful control performance limit DRL's adaptability and practicality in building control applications. To address these issues, this study introduces a new method, called Generative Adversarial Imitation Learning (GAIL), which effectively utilizes expert knowledge and demonstrations. Expert demonstrations range from fine-tuned rule-based controls to strategies inspired by optimization algorithms. By combining the capabilities of the generative adversarial network and imitation learning, GAIL is known for effectively learning the optimal strategy from expert demonstrations through an adversarial training process. We conducted a comprehensive evaluation comparing GAIL's performance with the DRL algorithm Proximal Policy Optimization (PPO) in the scenario of controlling a variable air volume system for load shifting in commercial buildings. Impressively, GAIL, guided by expert demonstrations based on model predictive control, achieved significantly improved computational efficiency and effectiveness. In terms of unified cumulative reward, GAIL with data augmentation achieved 95% expert performance, 22% higher than baseline rule-based control, in 100 training epochs; GAIL also outperformed PPO by 7%, resulting in 2% lower energy costs and notably improved thermal comfort. This improvement in thermal comfort is evidenced by a reduction of 18.65 unmet degree hours during the one-week operation. In comparison, PPO requires more training time and still lags behind GAIL in cumulative reward even after 500 epochs. These findings highlight the advantages of GAIL in enabling faster learning with fewer training samples, resulting in cost-effective solutions due to lower computational requirements. Overall, GAIL presents a promising approach to building energy management and provides a practical and flexible solution to the shortcomings of learning-based controllers that require extensive computational resources and training time.

DOI： 10.1016/j.apenergy.2024.123753

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Energy  

The prediction of renewable energy plays a critical role in optimizing the operation, fault diagnosis, and other essential tasks within its energy system. Given the scarcity of labeled data and the proliferation of newly established renewable energy systems, the concept of deep transfer learning can enhance the performance of deep learning prediction models in the renewable energy domain. Most existing studies primarily discuss the application of individual transfer learning algorithms, lacking comparative analysis and detailed methodological discourse among them. In this study, we compare the effectiveness of domain adaptation and fine-tuning as transfer learning methods in scenarios with limited labeled data. Furthermore, we introduce a composite transfer learning framework that initially applies domain adaptation followed by fine-tuning. Utilizing solar radiation data measured in Tokyo and Okinawa, we designed two sets of experiments with interchangeable source and target domains to verify the effectiveness and robustness of the proposed model. The experimental outcomes indicate that the sequential application of domain adaptation followed by fine-tuning surpasses the standalone use of either method, achieving prediction accuracy up to 98.89 % of the model trained with two full years of data. Additionally, this approach demonstrates superior prediction stability and lower outlier values.

DOI： 10.1016/j.energy.2024.131863

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

In the context of escalating energy consumption in buildings, particularly from air conditioning (AC), the intelligent control of AC has become increasingly crucial. Accurately predicting future energy consumption for AC, the indoor environment, and determining the optimal settings have emerged as key challenges in intelligent AC control. In this study, a hybrid self-learning dynamic graph neural network with self-attention mechanism is proposed for AC forecasting. Addressing the gaps in the existing graph neural network applications, this model overcomes the limitations of static graph structures by constructing evolving adjacency matrices integrated with a gated recurrent unit and self-attention, effectively capturing the dynamic relationships between changing feature quantities. Additionally, a multi-task prediction (MTP) module that utilizes both past and future data is proposed. The MTP enables the application of a single model to multiple prediction tasks, thereby obviating the need for separate model training for each task. An experiment in an actual outdoor environment was designed to verify the predictive performance of the proposed model. The results indicate that the proposed model achieves superior accuracy for all target variables across different tasks under various AC conditions, particularly for variables with strong non-linearity, which showed a maximum improvement of 24.94% in correlation coefficient (R2) compared to long-short term memory network. With the MTP, the single model applied to multiple prediction tasks exhibited only a minimal sacrifice in accuracy, resulting in a mere 0.64% decrease in average R2 of all target variables for the proposed model.

DOI： 10.1016/j.apenergy.2024.123156

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Building and Environment  

Control strategies for variable air volume (VAV) air-conditioning systems play a pivotal role in ensuring indoor environmental quality and energy efficiency. However, conventional approaches, such as static pressure reset (SPR) control, focus on managing indoor air temperature without considering the room pressure, which can lead to unbalanced room pressure and undesirable air leakage. Moreover, with the application of prevalent building pressure control strategies, such as airflow tracking control, to multizone VAV systems, neutralization of the room pressure is difficult across multiple zones in VAV systems. Therefore, this study introduces a model-based optimal control strategy for multizone VAV air-conditioning systems. The proposed strategy uses a multiobjective optimization framework to regulate fan frequencies and damper openings on both the supply and return sides. This holistic approach facilitates the simultaneous control of the indoor air temperature and room pressure while minimizing fan energy consumption. To assess the effectiveness of the proposed strategy, four control strategies were tested using a Python-based simulation testbed. The results demonstrate that the proposed strategy effectively maintains the indoor air temperature, neutralizes room pressure, and reduces fan energy consumption, thereby contributing to the overall efficiency of the VAV system. Moreover, the results highlight the limitations associated with combining airflow tracking control with SPR control for room pressure regulation in multizone VAV systems. This highlights the importance of adopting a model-based approach to address the complexities of concurrent room pressure and indoor air temperature control.

DOI： 10.1016/j.buildenv.2024.111464

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Accurate predictions of photovoltaic power generation (PV power) are essential for the integration of renewable energy into grids, markets, and building energy management systems. PV power is highly susceptible to weather conditions. Therefore, as weather forecast accuracy improves, it has become increasingly important issue to effectively utilize weather forecast data to enhance prediction accuracy. In this study, an improved model that combines Long Short-Term Memory (LSTM) and self-attention mechanisms is proposed. Proposed model captures the time features through the LSTM network and the correlations among multivariate time series through the self-attention mechanism. Additionally, methods to efficiently integrate historical and forecast data into various time-series forecasting models are also proposed. To verify the effectiveness of the proposed method and the performance of the proposed model, an actual PV power data of a building in Japan is used for various types of experiments. The results demonstrate that the proposed method effectively leverages weather forecast data and enhances the prediction performance of all models, the coefficient of determination (R2) are improved 15.8% for LSTM model, and 26.4% for proposed model. Whether for short-term or long-term predictions, proposed model consistently provides superior accuracy, practicality, and adaptability across all output sequence lengths. Compared to the basic LSTM model, R2 on short-term and long-term forecasting increased by 3.9% and 22.5%, respectively.

DOI： 10.1016/j.apenergy.2024.122709

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Applied Energy  

Deep learning models are increasingly applied in the field of solar radiation prediction. However, the substantial demand for labeled data limits their rapid application in newly established systems. Traditional transfer learning employs pre-training and fine-tuning methods to reduce the use of data in the target system. However, it still necessitates a small amount of labeled data for fine-tuning. This results in extensive time and cost for data collection, delaying the deployment of prediction models and optimization algorithms and leading to energy wastage. In this study, we employed the Adversarial Discriminative Domain Adaptation (ADDA) approach to achieve transfer learning under zero-label conditions in the target system, enabling new systems to harness the knowledge from other systems to create predictive models. Using the measured solar radiation data from Tokyo and Okinawa, two sets of experiments were designed with interchanged source and target domains to validate the efficacy and robustness of the proposed model. The results indicate that compared with the method of directly using the source domain model, transfer learning can enhance the predictive accuracy of the test set by at least 14% in both experiments, exhibiting more stable predictive performance and reduced prediction outliers.

DOI： 10.1016/j.apenergy.2024.122685

Scopus

Publishing type：Research paper (scientific journal)   Publisher：Energy  

Reinforcement Learning (RL), a promising algorithm for the operational control of Heating, Ventilation, and Air Conditioning (HVAC) systems, has garnered considerable attention and applications. However, traditional RL algorithms typically do not incorporate predictive information for future scenarios, and only a limited number of studies have examined the enhancement and impact of predictive information on RL algorithms. To address the issue of coupling RL and predictive information in HVAC system operation optimization, we employed an open-source framework to examine the impact of various predictive information strategies on RL outcomes. We propose a joint gated recurrent unit (GRU)-RL algorithm to handle situations where a time-series exists in state space. The results from four classic test cases demonstrate that the proposed GRU-RL method can reduce operating costs by approximately 14.5% and increase comfort performance by 88.4% in indoor comfort control and cost-management tasks. Moreover, the GRU-RL method outperformed the conventional DRL method and was merely augmented with prediction information. In indoor temperature regulation, the GRU-RL algorithm improves control efficacy by 14.2% compared to models without predictive information and offers an approximately 5% improvement over traditional network models. Finally, all models were made open source for easy replication and further research.

DOI： 10.1016/j.energy.2024.130344

Scopus

Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.jobe.2023.107814

Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.buildenv.2023.110885

Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.apenergy.2023.121591

Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.enbuild.2023.113461

Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.apenergy.2023.121106

Publishing type：Research paper (scientific journal)   Publisher：Elsevier BV  

DOI： 10.1016/j.ijepes.2023.109024

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.jobe.2022.105682

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.apenergy.2022.119783

Scopus

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.apenergy.2022.120021

Scopus

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.scs.2022.104317

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.renene.2022.02.051

Scopus

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.apenergy.2022.119288

Scopus

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.enbuild.2021.111379

Scopus

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.enbuild.2020.110156

Scopus

Publishing type：Research paper (scientific journal)   Publisher：{IOP} Publishing  

DOI： 10.1088/1755-1315/294/1/012033

Publishing type：Research paper (scientific journal)   Publisher：{IOP} Publishing  

<jats:title>Abstract</jats:title>
<jats:p>In order to improve the operation, detection and diagnosis of district energy systems, it is necessary to develop energy demand prediction models. Several models for energy prediction have been proposed, including machine learning methods and time series analysis methods. Data-driven machine learning methods fail to achieve the expected accuracy due to the lack of measurement data and the uncertainty of weather forecasts, additionally it is not easy to obtain complete and long-term weather data sets of building as input data in China. In this case, a WT-ARIMA prediction model that combines wavelet transform and time series analysis without meteorological parameters can be a better choice. The predicted performance of the commonly used time series model, WT-ARIMA model and LSTM model was tested based on the energy consumption data for one year. The results show that the model proposed in this paper has a 20% accuracy improvement over the ARIMA model and can reduce data requirement with good forecasting accuracy compared with LSTM-h.</jats:p>

DOI： 10.1088/1755-1315/294/1/012031

Language：English   Publishing type：Research paper (scientific journal)   Publisher：Elsevier {BV}  

DOI： 10.1016/j.egypro.2018.09.056