Buildings account for 40% of global energy consumption, with 36% attributed to heating, ventilation, and air-conditioning (HVAC) systems. Therefore, optimizing thermal system management in buildings is crucial for a sustainable future. Data-driven control strategies like deep reinforcement learning (DRL) and model predictive control (MPC) have already shown great promise, but they depend on simulation models. However, these models oversimplify real-world dynamics, leading to a "sim-to-real" gap that impairs the performance of pretrained controllers. This paper aims to close this gap using an innovative hybrid approach. To close the sim-to-real gap, data-driven methods such as artificial neural networks (ANN) have shown promise, but unlike physics-based models, these models lack transparency, making it difficult to generalize across different HVAC systems and diagnose the sim-to-real gap. This paper proposes a hybrid approach, where data-driven models are added to a physics-based model on the level of the components, thus maintaining the physics-based model's explainability. Given the scarcity of data on HVAC systems with degenerating components, a simulated environment is developed to replace measured data, meaning that in fact a sim-to-sim gap is closed. Using this environment, the methodology is validated for different small use cases by means of root mean squared error. It was found that the system's parameters are not required to develop the ANNs. Finally, it was observed that the number of neurons in the first hidden layer of the model is linked to that component behaving differently, suggesting a fault detection strategy based on the ANN's architecture.

ABSTRACT This study provides a comprehensive evaluation for the prediction of wind power ramping events in the Belgian Offshore Zone. These rapid, large‐scale power fluctuations pose significant challenges to grid reliability. The research uses operational Numerical Weather Prediction (NWP) models from the Royal Meteorological Institute of Belgium, as well as its version enhanced with Wind Farm Parameterization (WFP). Power predictions are generated with both typical power curves and machine learning approaches. Standard verification metrics, such as Mean Absolute Error (MAE), often fail to capture the operational significance of ramp events. To address this, we develop a flexible verification framework designed to assess ramp forecast performance. This framework incorporates adjustable time and power buffers, which tolerate minor, operationally acceptable discrepancies in the timing and magnitude of predicted events. Application of this framework to both intraday and day‐ahead forecasts reveals that WFP‐enhanced models consistently improve ramp predictions over the operational baseline. Further analysis reveals that while the WFP model with power curves effectively reduced false alarms, it comes at the cost of more misses. In contrast, ML‐based approaches achieve slightly higher overall skill scores by striking a better balance between reducing these error types. Moreover, we introduce the Ramp Alignment Score (RAS), an event‐based metric that quantifies the temporal alignment between predicted and observed ramps, to supplement the model evaluation by lead time. RAS analysis demonstrates that WFP models achieve better temporal alignment and reveals a distinct diurnal cycle in ramping prediction errors. Finally, we investigate the impact of a specific meteorological driver, finding an association between severe precipitation and large, highly predictable ramp events. Conversely, moderate and light precipitation are linked to a higher incidence of missed events and false alarms. This work provides both an operationally relevant evaluation methodology and insights into ramp predictions under specific meteorological conditions.

Accurate fault detection in district heating and cooling systems remains a challenge due to limitations in traditional simulation models. This paper explores a hybrid modelling approach that combines physics-based models with artificial neural networks to address this issue. Two methods of integrating these models are tested on simulated scenarios that replicate complex component behaviour. The first method applies corrections after full system simulation, while the second corrects outputs at the component level in real time. Both approaches reduce the discrepancy between simulated and reference data, but only the second method shows clear architectural changes based on component disturbances. This suggests a stronger potential for identifying faulty behaviour without the need for labelled datasets. The results demonstrate the value of integrating data-driven corrections at the component level to improve simulation accuracy and support fault detection in district heating and cooling systems.

Communication is crucial in multi-agent reinforcement learning when agents are not able to observe the full state of the environment. The most common approach to allow learned communication between agents is the use of a differentiable communication channel that allows gradients to flow between agents as a form of feedback. However, this is challenging when we want to use discrete messages to reduce the message size, since gradients cannot flow through a discrete communication channel. Previous work proposed methods to deal with this problem. However, these methods are tested in different communication learning architectures and environments, making it hard to compare them. In this paper, we compare several state-of-the-art discretization methods as well as a novel approach. We do this comparison in the context of communication learning using gradients from other agents and perform tests on several environments. In addition, we present COMA-DIAL, a communication learning approach based on DIAL and COMA extended with learning rate scaling and adapted exploration. Using COMA-DIAL allows us to perform experiments on more complex environments. Our results show that the novel ST-DRU method, proposed in this paper, achieves the best results out of all discretization methods across the different environments. It achieves the best or close to the best performance in each of the experiments and is the only method that does not fail on any of the tested environments.

Heat pump-based combined heat distribution circuits (CHDC) deliver both domestic hot water (DHW) and space heating to different end-users within an apartment building. By intelligently controlling heat pumps based on variable price signals, CHDCs can enhance sustainable energy use while reducing operational costs. Moreover, integrating decentralized DHW storage tanks enable decoupling of production and demand, leading to effective demand-response control strategies. This research examines the balance between thermal comfort and electricity costs by investigating the effectiveness of different supply temperature control strategies for various boundary conditions. Key challenges include maximizing thermal energy stored at DHW temperatures during low electricity price periods, while avoiding high-price periods, and ensuring end-users thermal comfort. Besides two baseline strategies, six demand-response control strategies, derived from a grouped charging methodology that alternates between high and low supply temperatures to reduce overall energy use, are investigated. Using two time-varying day-ahead marked price signals, i.e. stable and volatile, we determine the optimal control strategies for four different family types with different DHW consumption patterns. The results highlight that hybrid strategies reduce costs by up to 30

Physics-based ensemble weather prediction models form the backbone of probabilistic operational weather forecasting systems. The ensemble weather forecasts, however, suffer from systematic biases and inappropriate dispersion, which can be corrected by applying statistical postprocessing techniques. In this work, we apply a Transformer based on self-attention to postprocess forecasts of solar radiation from the EUPPBenchmark dataset. Our method results in a 5.3% improvement in CRPS over the raw forecasts, realizes a significant increase of 25 % in ensemble spread, and outperforms a classical member-by-member method employed as a competitive baseline. Finally, we convert these postprocessed weather forecasts into solar power predictions, highlighting the potential of the Transformer for practical renewable energy applications.

The increasing penetration of renewable energy sources and the electrification of residential load have amplified volatility in distribution networks. This paper investigates the influence of human behavior on household electricity usage and investigates how smartphone-collected data can enhance single-household consumption forecasting. By incorporating features related to human activity into a state-of-the-art sequence-to-sequence model based on bidirectional long short-term memory (LSTM) with attention, prediction accuracy improves by 15–22

As renewable energy sources continue to account for an increasing proportion of Belgium's energy production, decision making in renewable energy production increasingly relies on accurate numerical weather prediction forecasts. For general applications, forecast validation often focuses on direct comparisons to observations for the whole domains of interest, while in this study we assess model performance specifically related to renewable energy productions. We perform extended verification of relevant variables (wind speed, temperature, solar radiation, etc.) from multiple high-resolution deterministic and ensemble weather forecast models operated in Belgium for the period of May 2021 - June 2023. The forecasts are verified with observational datasets collected from on- and offshore weather stations, masts, lidars, and wind farm observations to comprehensively understand the capabilities of the models, making use of various deterministic and probabilistic skill scores. The results show that during lead times up to two days, although verification metrics differ among models, there are systematic errors in their forecasts for different observation sites. Such errors can often be eliminated by post-processing techniques. Therefore, we extend our verification dataset, with post-processed forecasts corrected by several methods including member-by-member and AI-based approaches. The results of this work will lead to an enhanced understanding of current forecasting skills of the operational models, help to evaluate the effectiveness of goal-oriented post-processing methods, and provide a reference for Belgian sustainable energy stakeholders.

We present a model-based method to disaggregate residential smart meter data into behind-the-meter (BTM) production and consumption profiles without requiring PV measurements. The approach fits a PV system model to feed-in power under clear-sky conditions, minimizing ramp-period loss while enforcing peak-time validity constraints, and leverages high-resolution solar radiation data to reconstruct accurate production and consumption profiles. Evaluation on 33 households from the Pecan Street dataset yields normalized mean absolute errors of 0.038 for production and 0.031 for consumption relative to ground truth. The proposed method is non-intrusive and explainable, enabling smart meter data applications for grid management, balancing and demand-side flexibility.

Indirect models for renewable energy forecasting rely heavily on accurate weather predictions. Operational weather forecasting today is mainly based on numerical weather prediction models, often employing ensembles to estimate the day-to-day forecast uncertainty. To correct for errors due to simplifications in these models, inaccurate initial conditions, and representativeness problems, statistical postprocessing becomes necessary for these ensemble forecasts. Current postprocessing techniques often disregard possible inter-ensemble relationships by correcting each member separately, or employ a distributional approach that requires extra multivariate methods to restore spatio-temporal and inter-variable correlations. In this work, we tackle these shortcomings with an innovative, attention-based member-by-member approach which postprocesses each member individually while simultaneously integrating information from other ensemble members. Variables required for renewable energy forecasting are postprocessed at the station level by regressing ensemble forecasts of multiple predictors, including the forecasted variable itself, against observational data. The training data utilized is sourced from the EUPPBench dataset, which contains ensemble predictions from the integrated forecasting system of the ECMWF and corresponding observations. Transformer modules built around Self-Attention are employed to capture dependencies between different predictors, such as temperature and total cloud cover, next to significant relationships between the ensemble members themselves. Additionally, our model postprocesses the forecasts for all lead times simultaneously, taking into account the correlation between the postprocessed variable and forecasts generated at earlier and later lead times. This results in postprocessing techniques that can be employed in downstream applications for conversion to renewable energy forecasts.

Deep reinforcement learning (DRL) can be used to optimise the performance of Collective Heating Systems (CHS) by reducing operational costs while ensuring thermal comfort. However, heating systems often exhibit slow responsiveness to control inputs due to thermal inertia, which delays the effects of actions such as adapting temperature set points. This delayed feedback complicates the learning process for DRL agents, as it becomes more difficult to associate specific control actions with their outcomes. To address this challenge, this study evaluates four hyperparameter schemes during training. The focus lies on schemes with varying learning rate (the rate at which weights in neural networks are adapted) and/or discount factor (the importance the DRL agent attaches to future rewards). In this respect, we introduce the GALER approach, which combines the progressive increase of the discount factor with the reduction of the learning rate throughout the training process. The effectiveness of the four learning schemes is evaluated using the actor-critic Proximal Policy Optimization (PPO) algorithm for three types of CHS with a multi-objective reward function balancing thermal comfort and energy use or operational costs. The results demonstrate that energy-based reward functions allow for limited optimisation possibilities, while the GALER scheme yields the highest potential for price-based optimisation across all considered concepts. It achieved a 3%-15% performance improvement over other successful training schemes. DRL agents trained with GALER schemes strategically anticipate on high-price times by lowering the supply temperature and vice versa. This research highlights the advantage of varying both learning rates and discount factors when training DRL agents to operate in complex multi-objective environments with slow responsiveness.

Load forecasting plays a pivotal role in industrial demand response, enabling businesses to plan their electricity needs ahead of time through day-ahead scheduling. However, data is often limited or outdated due to frequent infrastructure modifications. To this end, load forecasting using limited data has recently attracted research interest. This paper explores the application of learning-based algorithms to day-ahead load forecasting in data-constrained environments. Moreover, we introduce Diff-Ensemble, an ensemble incorporating the long short-term memory (LSTM) and diffusion model, to enhance load forecasting capabilities when data is limited, and evaluate it using real-world data from an industrial site as part of the InStaFlex project. Results show that Diff-Ensemble reduces the normalized MAE (NMAE) by 8.1

The conventional approach for controlling the supply temperature in collective space heating networks relies on a predefined heating curve determined by outdoor temperature and heat emitter type. This prioritizes thermal comfort but lacks energetic and financial optimization. This research proposes an adaptive supply temperature control in well-insulated dwellings, responsive to diverse environmental parameters. The approach considers variable electricity prices and accommodates different indoor temperature set points in dwellings. The study evaluates the effectiveness of two Deep Reinforcement Learning (DRL) algorithms, i.e. Proximal Policy Optimization (PPO) and Deep Q-Network (DQN), across various scenarios. Results reveal that DQN excels in collective space heating systems with underfloor heating in each dwelling, while PPO proves superior for radiator-based systems. Both outperform the traditional heating curve, achieving up to 13.77% (DQN) and 16.15% (PPO) cost reduction while guaranteeing thermal comfort. Additionally, the research highlights the capability of DRL-based methods to dynamically set the supply temperature based on a cloud of set points, showcasing adaptability to diverse environmental factors and addressing the growing significance of indoor heat gains in well-insulated dwellings. This innovative approach holds promise for more efficient and environmentally conscious heating strategies within collective space heating networks.

Evaluations are presented for the prediction of wind power ramping events in the Belgian Offshore Zone. Two models from the Royal Meteorological Institute of Belgium are verified: the operational ALARO-4km and its version with Wind Farm Parameterization (WFP). Power predictions are produced using power curves and machine learning (ML). As standard metrics such as MAE are insufficient for evaluating ramps, the proposed framework incorporates time and power buffers, enabling a flexible assessment that tolerates minor errors. Results indicate that WFP models enhance ramping prediction skill, while ML provides more balanced forecasts by reducing both misses and false alarms. A Ramp Alignment Score is also introduced to quantify temporal errors by forecast lead time, confirming that WFP models yield smaller average timing errors. Moreover, the framework reveals that severe precipitation is a strong indicator of large, predictable ramps, whereas lighter precipitation is associated with greater forecast errors.

In this paper the use of graph retrieval augmented generation (GRAG) in domain-specific knowledge bases (DSKBs) which contain both public and private data is explored. Our proposed methodology utilizes a modified embedding-based GRAG implementation capable of preserving private information in DSKBs. Our privacy-preserving GRAG methodology is evaluated against a baseline GRAG implementation using a modified version of the MultiHopRAG dataset and based on three metrics: privacy preservation, data quality and computational performance. This research proves that our methodology outperforms the baseline and is capable of preserving private information in DSKBs without compromising on output quality and computational performance.

Learning to communicate in order to share state information is an active problem in the area of multi-agent reinforcement learning. The credit assignment problem, the non-stationarity of the communication environment and the problem of encouraging the agents to be influenced by incoming messages are major challenges within this research field which need to be overcome in order to learn a valid communication protocol. This paper introduces the novel multi-agent counterfactual communication learning (MACC) method which adapts counterfactual reasoning in order to overcome the credit assignment problem for communicating agents. Next, the non-stationarity of the communication environment, while learning the communication Q-function, is overcome by creating the communication Q-function using the action policy of the other agents and the Q-function of the action environment. As the exact method to create the communication Q-function can be computationally intensive for a large number of agents, two approximation methods are proposed. Additionally, a social loss function is introduced in order to create influenceable agents, which is required to learn a valid communication protocol. Our experiments show that MACC is able to outperform the state-of-the-art baselines in four different scenarios in the particle environment. Finally, we demonstrate the scalability of MACC in a matrix environment.

Current postprocessing techniques often require separate models for each lead time and disregard possible inter-ensemble relationships by either correcting each member separately or by employing distributional approaches. In this work, we tackle these shortcomings with an innovative, fast and accurate Transformer which postprocesses each ensemble member individually while allowing information exchange across variables, spatial dimensions and lead times by means of multi-headed self-attention. Weather forecasts are postprocessed over 20 lead times simultaneously while including up to fifteen meteorological predictors. We use the EUPPBench dataset for training which contains ensemble predictions from the European Center for Medium-range Weather Forecasts' integrated forecasting system alongside corresponding observations. The work presented here is the first to postprocess the ten and one hundred-meter wind speed forecasts within this benchmark dataset, while also correcting two-meter temperature. Our approach significantly improves the original forecasts, as measured by the CRPS, with 16.5% for two-meter temperature, 10% for ten-meter wind speed and 9% for one hundred-meter wind speed, outperforming a classical member-by-member approach employed as a competitive benchmark. Furthermore, being up to six times faster, it fulfills the demand for rapid operational weather forecasts in various downstream applications, including renewable energy forecasting.

This study analyses the in-situ performance of Combined Heat Distribution Circuit systems and Heat Interface Units to address commonly overlooked faults. Using data from energy meters, including gas meters and automated heat meter readings from seven case studies in Antwerp (Belgium), we investigate the CHDC's overall efficiency, comparing it to expected values reported in the literature. Key performance indicators such as the overflow and VWART are determined to identify suboptimal performing HIUs. An in-depth analysis involves manual investigation of data and site visits to identify deviations indicating potential faults within the system. Our findings highlight the significant impact of faults in the hydronic circuit on overall system performance, emphasizing the need to transition towards data-driven techniques and machine learning methods for fault detection in collective heat systems.

Automated systems are increasingly integrated into our daily lives, streamlining various tasks and enhancing convenience. Despite their careful design to improve our everyday experiences, problems still occur, often due to human interaction with these systems. This paper addresses the challenges posed by human impact on autonomous systems, aiming to predict and mitigate errors caused by such interactions. By incorporating human behavior into the training process, we hypothesize that the trained agent’s ability to anticipate and withstand these behaviors will improve. Leveraging artificial intelligence (AI) and reinforcement learning (RL) in particular, a controller is developed for automated processes designed to anticipate human impact and minimize errors. We differentiate irrational human behavior into two categories: short-term irrationality and long-term irrationality. This research focuses on the short-term irrational behaviors as a manageable subset. To address the lack of data on irrational human behavior, we define an irrational model within a straightforward environment to evaluate RL’s ability to recognize and anticipate such behavior. The environment used is the card game UNO, because of its simple rules and emphasis on player interaction. Results reveal a notable 0.4