Pub Date : 2025-11-01DOI: 10.1016/j.egyai.2025.100639
Joseph Collins, Andreas Amann, Kieran Mulchrone
Fundamental short-term electricity market models typically rely on expert-driven assumptions and manual, iterative calibration, and often neglect strategic bidding behaviours. To address these issues, we develop a bottom-up (fundamental) model that forecasts day-ahead outcomes from publicly available participant order data using regularised regression, machine learning, and neural network techniques. We introduce an explainability framework that decomposes forecast errors at participant, cohort, and aggregate levels, linking forecast performance to forecast trading behaviours. Compared with a benchmark top-down model, the bottom-up approach yields lower price forecast accuracy but demonstrates an ability to capture market dynamics. Where forecast dynamics diverge from observed outcomes, many misaligned cases are attributable to specific cohorts, particularly financial traders (speculators). Beyond forecasting, the framework offers complementary applications: the modelling approach can support calibration of traditional fundamental models and serve as a stand-alone forecaster in markets beyond day-ahead where order data are available, while the explainability component can apply to both bottom-up and hybrid modelling approaches. The study highlights the challenges inherent in bottom-up fundamental models, while showing how our approach provides new insights and practical tools to support their calibration and application.
{"title":"A new approach to fundamental electricity market modelling: Exploring market dynamics and speculator influence","authors":"Joseph Collins, Andreas Amann, Kieran Mulchrone","doi":"10.1016/j.egyai.2025.100639","DOIUrl":"10.1016/j.egyai.2025.100639","url":null,"abstract":"<div><div>Fundamental short-term electricity market models typically rely on expert-driven assumptions and manual, iterative calibration, and often neglect strategic bidding behaviours. To address these issues, we develop a bottom-up (fundamental) model that forecasts day-ahead outcomes from publicly available participant order data using regularised regression, machine learning, and neural network techniques. We introduce an explainability framework that decomposes forecast errors at participant, cohort, and aggregate levels, linking forecast performance to forecast trading behaviours. Compared with a benchmark top-down model, the bottom-up approach yields lower price forecast accuracy but demonstrates an ability to capture market dynamics. Where forecast dynamics diverge from observed outcomes, many misaligned cases are attributable to specific cohorts, particularly financial traders (speculators). Beyond forecasting, the framework offers complementary applications: the modelling approach can support calibration of traditional fundamental models and serve as a stand-alone forecaster in markets beyond day-ahead where order data are available, while the explainability component can apply to both bottom-up and hybrid modelling approaches. The study highlights the challenges inherent in bottom-up fundamental models, while showing how our approach provides new insights and practical tools to support their calibration and application.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100639"},"PeriodicalIF":9.6,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145465941","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-31DOI: 10.1016/j.egyai.2025.100643
Moisés Cordeiro-Costas , Raquel Pérez-Orozco , Pablo Hernandez-Cruz , Francisco Troncoso-Pastoriza , Enrique Granada-Álvarez
Understanding occupancy patterns in buildings is critical for optimizing energy use, improving indoor comfort, and enabling smarter building management systems. Traditional methods for occupancy detection often rely on dense networks of sensors or invasive technologies such as cameras or wearables, raising concerns about cost, scalability, and privacy. In contrast, data-driven approaches based on environmental and energy-related signals offer a promising alternative: they are cost-effective, unobtrusive, and compatible with existing infrastructure. This study presents a robust and generalizable deep learning framework for occupancy estimation using easily accessible data sources, including electricity consumption, CO2 concentration, and working-hour schedules. Through an extensive comparative analysis of different input combinations in supervised model configurations, the optimal trade-off between accuracy and computational cost for real-time deployment is identified. The results highlight the value of selecting appropriate variables and reveal how models using minimal inputs can provide reliable estimations when properly designed. By demonstrating a non-invasive, privacy-preserving, and scalable approach to occupancy modeling, this work contributes to the development of energy-aware and intelligent buildings, essential for meeting sustainability goals and enhancing user-centric building automation.
{"title":"Hybrid LSTM-MLP model with NSGA-II-based hyperparameter optimization for non-invasive occupancy estimation","authors":"Moisés Cordeiro-Costas , Raquel Pérez-Orozco , Pablo Hernandez-Cruz , Francisco Troncoso-Pastoriza , Enrique Granada-Álvarez","doi":"10.1016/j.egyai.2025.100643","DOIUrl":"10.1016/j.egyai.2025.100643","url":null,"abstract":"<div><div>Understanding occupancy patterns in buildings is critical for optimizing energy use, improving indoor comfort, and enabling smarter building management systems. Traditional methods for occupancy detection often rely on dense networks of sensors or invasive technologies such as cameras or wearables, raising concerns about cost, scalability, and privacy. In contrast, data-driven approaches based on environmental and energy-related signals offer a promising alternative: they are cost-effective, unobtrusive, and compatible with existing infrastructure. This study presents a robust and generalizable deep learning framework for occupancy estimation using easily accessible data sources, including electricity consumption, CO<sub>2</sub> concentration, and working-hour schedules. Through an extensive comparative analysis of different input combinations in supervised model configurations, the optimal trade-off between accuracy and computational cost for real-time deployment is identified. The results highlight the value of selecting appropriate variables and reveal how models using minimal inputs can provide reliable estimations when properly designed. By demonstrating a non-invasive, privacy-preserving, and scalable approach to occupancy modeling, this work contributes to the development of energy-aware and intelligent buildings, essential for meeting sustainability goals and enhancing user-centric building automation.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100643"},"PeriodicalIF":9.6,"publicationDate":"2025-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145465939","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-29DOI: 10.1016/j.egyai.2025.100644
Yunfei Mu , Ruichao Zhou , Kangning Zhao , Hongjie Jia , Guoqiang Zu , Ye Yang
The charging behaviors of electric vehicle (EV) users exhibit high randomness and individual heterogeneity, with the key parameters such as the charging duration and charged energy levels displaying significant fluctuations. Compared with EV cluster-layer prediction, predicting the charging demands of individual users requires not only the analysis of more complex charging behaviors but also the establishment of a coupling model between exogenous variables (e.g., weather and holidays) and prediction accuracy, thereby imposing higher robustness requirements on prediction algorithms. An individual-user EV charging demand prediction method that integrates multisource data with a dual-layer clustering approach and a light gradient boosting machine (LightGBM) is proposed in this study to address these technical challenges. First, a multisource dataset that incorporates user charging behavior data and exogenous variables (meteorological factors and date types) is constructed. A dual-layer feature extraction mechanism consisting of data-layer clustering for charging type identification and user-layer clustering for user group classification is employed, thereby establishing a classification feature space that characterizes different charging types and user groups. A predictive model is subsequently developed using the LightGBM algorithm, which directly incorporates classification features as its inputs, effectively mitigating the information loss associated with the traditional categorical variable encoding process. Finally, employing EV users from a typical residential community in northern China as an empirical case, comparative experiments are performed to validate the proposed method, demonstrating its effectiveness at improving prediction accuracy.
{"title":"A charging demand prediction method for individual electric vehicle users based on dual-layer multisource data clustering and a LightGBM","authors":"Yunfei Mu , Ruichao Zhou , Kangning Zhao , Hongjie Jia , Guoqiang Zu , Ye Yang","doi":"10.1016/j.egyai.2025.100644","DOIUrl":"10.1016/j.egyai.2025.100644","url":null,"abstract":"<div><div>The charging behaviors of electric vehicle (EV) users exhibit high randomness and individual heterogeneity, with the key parameters such as the charging duration and charged energy levels displaying significant fluctuations. Compared with EV cluster-layer prediction, predicting the charging demands of individual users requires not only the analysis of more complex charging behaviors but also the establishment of a coupling model between exogenous variables (e.g., weather and holidays) and prediction accuracy, thereby imposing higher robustness requirements on prediction algorithms. An individual-user EV charging demand prediction method that integrates multisource data with a dual-layer clustering approach and a light gradient boosting machine (LightGBM) is proposed in this study to address these technical challenges. First, a multisource dataset that incorporates user charging behavior data and exogenous variables (meteorological factors and date types) is constructed. A dual-layer feature extraction mechanism consisting of data-layer clustering for charging type identification and user-layer clustering for user group classification is employed, thereby establishing a classification feature space that characterizes different charging types and user groups. A predictive model is subsequently developed using the LightGBM algorithm, which directly incorporates classification features as its inputs, effectively mitigating the information loss associated with the traditional categorical variable encoding process. Finally, employing EV users from a typical residential community in northern China as an empirical case, comparative experiments are performed to validate the proposed method, demonstrating its effectiveness at improving prediction accuracy.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100644"},"PeriodicalIF":9.6,"publicationDate":"2025-10-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145465937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-29DOI: 10.1016/j.egyai.2025.100641
Simone Eiraudo , Daniele Salvatore Schiera , Luca Barbierato , Alena Trifirò , Lorenzo Bottaccioli , Andrea Lanzini
An ecosystem of energy models of buildings is needed to boost the retrofitting process to improve energy efficiency and meet sustainability goals. Such models should enhance the understanding of the energy behavior of a building, the impact of the external variables, and the causes of inefficiencies. Energy Signatures can fill this role, with particular regard to the consumption due to air conditioning. Univariate models, neglecting the impact of solar radiation, have been widely adopted for Energy Signature analysis. This paper presents Multivariable Energy Signatures considering outdoor temperature and solar radiation. The application on a real-world dataset of multivariable non-parametric approaches stands out from previous works in the ES sector. This led to a mean improvement of 0.768 to 0.804 of the coefficients of determination calculated over 103 real-world case studies. Moreover, Neural Networks outperformed several literature algorithms regarding accuracy, robustness, and scalability. The paper also discusses issues regarding the time resolution of input data and introduces appropriate visualization tools to employ Multivariable Energy Signatures as diagnostic tools.
{"title":"A comparative analysis of regression algorithms and a real world application of multivariable energy signatures","authors":"Simone Eiraudo , Daniele Salvatore Schiera , Luca Barbierato , Alena Trifirò , Lorenzo Bottaccioli , Andrea Lanzini","doi":"10.1016/j.egyai.2025.100641","DOIUrl":"10.1016/j.egyai.2025.100641","url":null,"abstract":"<div><div>An ecosystem of energy models of buildings is needed to boost the retrofitting process to improve energy efficiency and meet sustainability goals. Such models should enhance the understanding of the energy behavior of a building, the impact of the external variables, and the causes of inefficiencies. Energy Signatures can fill this role, with particular regard to the consumption due to air conditioning. Univariate models, neglecting the impact of solar radiation, have been widely adopted for Energy Signature analysis. This paper presents Multivariable Energy Signatures considering outdoor temperature and solar radiation. The application on a real-world dataset of multivariable non-parametric approaches stands out from previous works in the ES sector. This led to a mean improvement of 0.768 to 0.804 of the coefficients of determination calculated over 103 real-world case studies. Moreover, Neural Networks outperformed several literature algorithms regarding accuracy, robustness, and scalability. The paper also discusses issues regarding the time resolution of input data and introduces appropriate visualization tools to employ Multivariable Energy Signatures as diagnostic tools.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100641"},"PeriodicalIF":9.6,"publicationDate":"2025-10-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145465936","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-28DOI: 10.1016/j.egyai.2025.100635
Haoyang Zhang , Georgios Tsaousoglou , Sen Zhan , Koen Kok , Nikolaos G. Paterakis
Distributed energy resources, connected to power distribution systems, are increasingly operated by intelligent/learning agents. Such agents, looking to optimize their own payoff, can discover harmful ways to exploit the system. Hence, shielding critical systems of harmful agent behavior is of crucial importance. In this paper, the problem of designing an efficient operating mechanism for a power distribution system is taken on, considering the realistic case where the system’s resources do not possess this information and instead learn to improve their policies through experience. To that end, a multi-agent reinforcement learning algorithm is developed to model the participants’ learning-to-act process and consider the agents’ learning under different pricing schemes that shape the agents’ reward functions. Two popular pricing schemes (pay-as-bid and distribution locational marginal pricing) are presented, exposing that learning agents can discover ways to exploit them, resulting in severe dispatch inefficiency. A game-theoretic pricing scheme is presented that theoretically incentivizes truthful agent behavior, and empirically demonstrate that this property improves the efficiency of the resulting dispatch also in the presence of learning agents. In particular, the proposed scheme is able to outperform the popular distribution locational marginal pricing scheme, in terms of efficiency, by a factor of 15–17%.
{"title":"Taming deep reinforcement learning agents with pricing mechanism: Validation in power distribution systems","authors":"Haoyang Zhang , Georgios Tsaousoglou , Sen Zhan , Koen Kok , Nikolaos G. Paterakis","doi":"10.1016/j.egyai.2025.100635","DOIUrl":"10.1016/j.egyai.2025.100635","url":null,"abstract":"<div><div>Distributed energy resources, connected to power distribution systems, are increasingly operated by intelligent/learning agents. Such agents, looking to optimize their own payoff, can discover harmful ways to exploit the system. Hence, shielding critical systems of harmful agent behavior is of crucial importance. In this paper, the problem of designing an efficient operating mechanism for a power distribution system is taken on, considering the realistic case where the system’s resources do not possess this information and instead learn to improve their policies through experience. To that end, a multi-agent reinforcement learning algorithm is developed to model the participants’ learning-to-act process and consider the agents’ learning under different pricing schemes that shape the agents’ reward functions. Two popular pricing schemes (pay-as-bid and distribution locational marginal pricing) are presented, exposing that learning agents can discover ways to exploit them, resulting in severe dispatch inefficiency. A game-theoretic pricing scheme is presented that theoretically incentivizes truthful agent behavior, and empirically demonstrate that this property improves the efficiency of the resulting dispatch also in the presence of learning agents. In particular, the proposed scheme is able to outperform the popular distribution locational marginal pricing scheme, in terms of efficiency, by a factor of 15–17%.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100635"},"PeriodicalIF":9.6,"publicationDate":"2025-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416161","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Accurately predicting and optimizing heating and cooling demands in building energy management is crucial for enhancing efficiency and reducing energy consumption. Traditional methods often struggle with building energy usage patterns' nonlinear and variable nature. With the advent of advanced data collection through smart sensors, there is a growing need for intelligent systems to leverage this data to provide actionable insights. This study addresses the gap by developing a recommendation system using three machine and deep learning models, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and XGBoost to predict and optimize the efficiency levels of space heating, ceiling cooling, and free cooling systems. Our proposed solution harnesses the power of these models, with LSTM performing best overall, to forecast energy consumption across hourly and daily timescales, enabling precise adjustments and efficient energy management. The methodology involves extensive data preprocessing, including hierarchical imputation of missing values and label encoding of categorical variables, followed by the transformation of raw data into efficiency levels. The deep learning model architecture, consisting of sequential layers, captures long-term dependencies in the data, while grid search-based hyperparameter tuning optimizes model performance. Results indicate high predictive accuracy, with R-squared values demonstrating the model's ability to explain up to 97.2 % of the variance in hourly space heating, 95.2 % in daily ceiling cooling, and 93 % in daily free cooling energy consumption. Additionally, we interpret graphs using OpenAI's GPT-4 model to enhance understanding and facilitate actionable insights. This interpretation enhances the clarity of the predictive results, supporting more informed decision-making in energy management. The significance of this work lies in its potential to transform energy management practices in building environments, providing a robust tool for optimizing heating and cooling operations and contributing to overall energy efficiency.
{"title":"Smart energy strategies: Leveraging LSTM and LLMs for advanced energy management","authors":"Fernando Almeida , Mauro Castelli , Nadine Côrte-Real , Camilla Fallarino , Luca Manzoni","doi":"10.1016/j.egyai.2025.100642","DOIUrl":"10.1016/j.egyai.2025.100642","url":null,"abstract":"<div><div>Accurately predicting and optimizing heating and cooling demands in building energy management is crucial for enhancing efficiency and reducing energy consumption. Traditional methods often struggle with building energy usage patterns' nonlinear and variable nature. With the advent of advanced data collection through smart sensors, there is a growing need for intelligent systems to leverage this data to provide actionable insights. This study addresses the gap by developing a recommendation system using three machine and deep learning models, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and XGBoost to predict and optimize the efficiency levels of space heating, ceiling cooling, and free cooling systems. Our proposed solution harnesses the power of these models, with LSTM performing best overall, to forecast energy consumption across hourly and daily timescales, enabling precise adjustments and efficient energy management. The methodology involves extensive data preprocessing, including hierarchical imputation of missing values and label encoding of categorical variables, followed by the transformation of raw data into efficiency levels. The deep learning model architecture, consisting of sequential layers, captures long-term dependencies in the data, while grid search-based hyperparameter tuning optimizes model performance. Results indicate high predictive accuracy, with R-squared values demonstrating the model's ability to explain up to 97.2 % of the variance in hourly space heating, 95.2 % in daily ceiling cooling, and 93 % in daily free cooling energy consumption. Additionally, we interpret graphs using OpenAI's GPT-4 model to enhance understanding and facilitate actionable insights. This interpretation enhances the clarity of the predictive results, supporting more informed decision-making in energy management. The significance of this work lies in its potential to transform energy management practices in building environments, providing a robust tool for optimizing heating and cooling operations and contributing to overall energy efficiency.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100642"},"PeriodicalIF":9.6,"publicationDate":"2025-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416080","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-28DOI: 10.1016/j.egyai.2025.100634
Muhammad Hafeez Saeed , Maomao Hu , Hussain Kazmi , Geert Deconinck
Delivering energy flexibility at the district scale entails coordinating control actions across many buildings to shape aggregate demand; this coordination depends on training and deploying control policies and optimization routines, which in turn require predictive models that can be queried efficiently over large building clusters. However, conventional physics-based simulators are computationally prohibitive for large-scale control training, and simple data-driven surrogates often lack the generalization needed for heterogeneous clusters. This paper introduces ScaleONet, a deep operator network framework designed for scalable, control-oriented modeling of building-cluster thermal dynamics. ScaleONet leverages the DeepONet paradigm to decouple and share learning across buildings: an LSTM-based branch network encodes outdoor climate and individual HVAC control signals, while a multilayer perceptron (MLP)-based trunk network embeds prediction timestamps, enabling fast predictions for growing clusters with negligible extra cost for each additional building or timestep. To the authors’ knowledge, this is the first operator-learning method tailored to indoor air temperature forecasting in heterogeneous building clusters. Validation on thirty Belgian buildings (GenkNet) simulated in Dymola shows that, although a non-operator-learning LSTM baseline slightly outperforms ScaleONet for single-building cases, its error grows monotonically with cluster size. In contrast, ScaleONet’s median per-building-per-day RMSE decreases from 0.59 °C at three buildings to 0.53 °C at ten and 0.47 °C at thirty, compared to 0.95 °C for the LSTM at thirty buildings — a 51% reduction in prediction error. Error analysis across envelope heat-loss coefficients () further reveals that while the LSTM’s RMSE increases for high- structures, ScaleONet maintains uniformly low error. With millisecond-scale inference (approximately per sample for thirty buildings), ScaleONet is well suited for large-scale reinforcement learning, receding-horizon optimization, and real-time model predictive control.
{"title":"ScaleONet: Scalable and control-oriented modeling of building cluster thermal dynamics using deep operator networks — A practical case study for a Belgian district","authors":"Muhammad Hafeez Saeed , Maomao Hu , Hussain Kazmi , Geert Deconinck","doi":"10.1016/j.egyai.2025.100634","DOIUrl":"10.1016/j.egyai.2025.100634","url":null,"abstract":"<div><div>Delivering energy flexibility at the district scale entails coordinating control actions across many buildings to shape aggregate demand; this coordination depends on training and deploying control policies and optimization routines, which in turn require predictive models that can be queried efficiently over large building clusters. However, conventional physics-based simulators are computationally prohibitive for large-scale control training, and simple data-driven surrogates often lack the generalization needed for heterogeneous clusters. This paper introduces <em>ScaleONet</em>, a deep operator network framework designed for scalable, control-oriented modeling of building-cluster thermal dynamics. ScaleONet leverages the DeepONet paradigm to decouple and share learning across buildings: an LSTM-based branch network encodes outdoor climate and individual HVAC control signals, while a multilayer perceptron (MLP)-based trunk network embeds prediction timestamps, enabling fast predictions for growing clusters with negligible extra cost for each additional building or timestep. To the authors’ knowledge, this is the first operator-learning method tailored to indoor air temperature forecasting in heterogeneous building clusters. Validation on thirty Belgian buildings (GenkNet) simulated in Dymola shows that, although a non-operator-learning LSTM baseline slightly outperforms ScaleONet for single-building cases, its error grows monotonically with cluster size. In contrast, ScaleONet’s median per-building-per-day RMSE decreases from 0.59 °C at three buildings to 0.53 °C at ten and 0.47 °C at thirty, compared to 0.95 °C for the LSTM at thirty buildings — a 51% reduction in prediction error. Error analysis across envelope heat-loss coefficients (<span><math><msub><mrow><mi>UA</mi></mrow><mrow><mi>building</mi></mrow></msub></math></span>) further reveals that while the LSTM’s RMSE increases for high-<span><math><mrow><mi>U</mi><mi>A</mi></mrow></math></span> structures, ScaleONet maintains uniformly low error. With millisecond-scale inference (approximately <span><math><mrow><mn>4</mn><mspace></mspace><mi>ms</mi></mrow></math></span> per sample for thirty buildings), ScaleONet is well suited for large-scale reinforcement learning, receding-horizon optimization, and real-time model predictive control.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100634"},"PeriodicalIF":9.6,"publicationDate":"2025-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416162","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-28DOI: 10.1016/j.egyai.2025.100637
Mengbo Yu, Alexander Neubauer, Pedram Babakhani, Stefan Brandt, Martin Kriegel
Accurately filling in missing heating data is essential for ensuring data quality in applications such as energy management optimization and building efficiency analysis. Traditional machine learning methods use historical heating data as an input feature to predict the following missing data. However, when the duration of missing data is long, previous estimated values are inevitably used for further imputation, leading to error accumulation and a growing deviation from true values. To overcome this problem, this paper proposes a generative network that can fill missing data solely based on weather and temporal data, without using previous imputed values for further imputation. Our method outperformed the state of the art such as Seq2seq and Transformer, achieving relative normalized root mean square error (NRMSE) reductions of 1.65 % to 41.38 %, 0.30 % to 66.43 %, and 14.84 % to 50.22 % across three different data sources. In addition, with our proposed method, the effect of selecting different weather variables on model performance, and the benefits of transfer learning under limited data were also demonstrated. The relative NRMSE reduction is between 3.88 % to 15.85 % in cold months and from 7.49 % to 12.29 % in warm months when applying transfer learning.
{"title":"Imputing the long-term missing heating load data using a generative network","authors":"Mengbo Yu, Alexander Neubauer, Pedram Babakhani, Stefan Brandt, Martin Kriegel","doi":"10.1016/j.egyai.2025.100637","DOIUrl":"10.1016/j.egyai.2025.100637","url":null,"abstract":"<div><div>Accurately filling in missing heating data is essential for ensuring data quality in applications such as energy management optimization and building efficiency analysis. Traditional machine learning methods use historical heating data as an input feature to predict the following missing data. However, when the duration of missing data is long, previous estimated values are inevitably used for further imputation, leading to error accumulation and a growing deviation from true values. To overcome this problem, this paper proposes a generative network that can fill missing data solely based on weather and temporal data, without using previous imputed values for further imputation. Our method outperformed the state of the art such as Seq2seq and Transformer, achieving relative normalized root mean square error (NRMSE) reductions of 1.65<!--> <!-->% to 41.38<!--> <!-->%, 0.30<!--> <!-->% to 66.43<!--> <!-->%, and 14.84<!--> <!-->% to 50.22<!--> <!-->% across three different data sources. In addition, with our proposed method, the effect of selecting different weather variables on model performance, and the benefits of transfer learning under limited data were also demonstrated. The relative NRMSE reduction is between 3.88<!--> <!-->% to 15.85<!--> <!-->% in cold months and from 7.49<!--> <!-->% to 12.29<!--> <!-->% in warm months when applying transfer learning.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100637"},"PeriodicalIF":9.6,"publicationDate":"2025-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416081","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-10-27DOI: 10.1016/j.egyai.2025.100633
Bizhi Wu , Shanlin Wang
The increasing integration of renewable energy sources introduces significant variability and non-stationarity into power system, challenging accurate net load forecasting. Although net load forecasting research has devoted considerable efforts to handle non-stationarity – via normalization, incremental learning, or drift detection – existing solutions often suffer from hyperparameter tuning, threshold-based triggers, or reliance on specialized architectures. To overcome these limitations, we propose Adaptive Smoothing Drift Normalization (ASDN), a lightweight normalization layer that continuously adapts to distribution shifts without threshold tuning. ASDN effectively adapts to new data via a mechanism that combines entropy-based adjustments with a dynamic filtering approach. At the same time, it maintains stability with respect to historical patterns, allowing the method to capture both gradual and abrupt shifts in the data distribution. We provide a theoretical guarantee that the estimation error of ASDN remains bounded under piecewise-stationary drift; as incremental drift and noise decrease, this bound tightens and converges to zero. Experiments on nine forecasting models across five public datasets and four prediction horizons show that ASDN consistently outperforms traditional normalization techniques, reducing mean squared error and enhancing robustness. These results confirm ASDN’s effectiveness in handling complex temporal dynamics, making it valuable for improving forecast accuracy in dynamic renewable power systems.
{"title":"Adaptive smoothing drift normalization for day-ahead net load forecasting in renewable power system","authors":"Bizhi Wu , Shanlin Wang","doi":"10.1016/j.egyai.2025.100633","DOIUrl":"10.1016/j.egyai.2025.100633","url":null,"abstract":"<div><div>The increasing integration of renewable energy sources introduces significant variability and non-stationarity into power system, challenging accurate net load forecasting. Although net load forecasting research has devoted considerable efforts to handle non-stationarity – via normalization, incremental learning, or drift detection – existing solutions often suffer from hyperparameter tuning, threshold-based triggers, or reliance on specialized architectures. To overcome these limitations, we propose Adaptive Smoothing Drift Normalization (ASDN), a lightweight normalization layer that continuously adapts to distribution shifts without threshold tuning. ASDN effectively adapts to new data via a mechanism that combines entropy-based adjustments with a dynamic filtering approach. At the same time, it maintains stability with respect to historical patterns, allowing the method to capture both gradual and abrupt shifts in the data distribution. We provide a theoretical guarantee that the estimation error of ASDN remains bounded under piecewise-stationary drift; as incremental drift and noise decrease, this bound tightens and converges to zero. Experiments on nine forecasting models across five public datasets and four prediction horizons show that ASDN consistently outperforms traditional normalization techniques, reducing mean squared error and enhancing robustness. These results confirm ASDN’s effectiveness in handling complex temporal dynamics, making it valuable for improving forecast accuracy in dynamic renewable power systems.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100633"},"PeriodicalIF":9.6,"publicationDate":"2025-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145465938","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Accurately estimating power loss reduction from passive filters before installation is challenging due to variable loads and power quality conditions across grid points. Existing studies rely on simulation or analytical models. These approaches often fail to capture real-world variability through data-driven methods. This gap limits effective, site-specific filter deployment decisions. We present a two-step machine learning approach to estimate energy efficiency gains from passive filters under variable conditions using high-resolution power analyzer data. Ridge Regression identifies key predictive variables, achieving baseline R² = 0.591. XGBoost then captures nonlinear interactions between load variability, power quality disturbances, and filter performance, improving accuracy to R² = 0.755. The methodology was validated through deployment at three industrial facilities in collaboration with Livarsa GmbH. Results demonstrate 9.9% average relative error across measured efficiency gains, confirming reliability under real-world conditions. Comprehensive validation through k-fold cross-validation, ensemble methods, and external testing quantified prediction uncertainty inherent in small industrial datasets (25 training samples). The approach offers a scalable, data-driven decision-support tool overcoming simulation-based limitations. Computational efficiency enables real-time assessment during client consultations without specialized software. Economic value derives from reduced performance guarantee margins, accelerated assessment timelines, and minimized warranty exposure. Limitations include statistical constraints from limited training data, reflected in cross-validation overfitting and wide confidence intervals. External validity requires site-specific validation for facilities with substantially different electrical characteristics. Despite these constraints, the findings provide practical value for energy professionals seeking efficient power quality solutions, enabling confident passive filter deployment decisions based on quantified performance predictions.
{"title":"Data-driven machine learning model estimates efficiency gains from passive filters under variable loads","authors":"Uchenna Johnpaul Aniekwensi , Dipyaman Basu , Jörg Bausch","doi":"10.1016/j.egyai.2025.100631","DOIUrl":"10.1016/j.egyai.2025.100631","url":null,"abstract":"<div><div>Accurately estimating power loss reduction from passive filters before installation is challenging due to variable loads and power quality conditions across grid points. Existing studies rely on simulation or analytical models. These approaches often fail to capture real-world variability through data-driven methods. This gap limits effective, site-specific filter deployment decisions. We present a two-step machine learning approach to estimate energy efficiency gains from passive filters under variable conditions using high-resolution power analyzer data. Ridge Regression identifies key predictive variables, achieving baseline R² = 0.591. XGBoost then captures nonlinear interactions between load variability, power quality disturbances, and filter performance, improving accuracy to R² = 0.755. The methodology was validated through deployment at three industrial facilities in collaboration with Livarsa GmbH. Results demonstrate 9.9% average relative error across measured efficiency gains, confirming reliability under real-world conditions. Comprehensive validation through k-fold cross-validation, ensemble methods, and external testing quantified prediction uncertainty inherent in small industrial datasets (25 training samples). The approach offers a scalable, data-driven decision-support tool overcoming simulation-based limitations. Computational efficiency enables real-time assessment during client consultations without specialized software. Economic value derives from reduced performance guarantee margins, accelerated assessment timelines, and minimized warranty exposure. Limitations include statistical constraints from limited training data, reflected in cross-validation overfitting and wide confidence intervals. External validity requires site-specific validation for facilities with substantially different electrical characteristics. Despite these constraints, the findings provide practical value for energy professionals seeking efficient power quality solutions, enabling confident passive filter deployment decisions based on quantified performance predictions.</div></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"22 ","pages":"Article 100631"},"PeriodicalIF":9.6,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145362268","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号