Peter Kepplinger

Accurate load forecasting is essential for the optimal operation of local energy communities. However, newly established communities usually lack sufficient historical data, limiting the use of supervised forecasting and optimization methods. This paper investigates the zero-shot time-series foundation model TiRex for weekly load forecasting in emerging energy communities. After an initial one-week observation period, the model is deployed without task-specific training, feature engineering, or exogenous input data. The approach is evaluated using ten months of real measurements from an energy community consisting of six residential households, one school, and one industrial building, including local photovoltaic generation. Using only historical load data, TiRex generates weekly forecasts in a walk-forward setup that reflects realistic operation. In addition to point forecasts, predictive quantiles are provided to represent forecast uncertainty. Performance is compared with a weekly persistence baseline, where each forecast week equals the load profile of the previous week. Despite requiring no training or external data, TiRex reduces forecasting error by about 20%, achieving a mean RMSE of 1.50 kW versus 1.89 kW for persistence. As an additional benchmark, the transformer-based foundation model Chronos is tested in three model sizes, reaching RMSE values between 1.59 kW and 1.65 kW under identical conditions. The results show that zero-shot forecasting can provide accurate and stable performance under realistic deployment conditions, even with distributed photovoltaic generation and volatile residual profiles. This enables forecasting-based operation from the earliest deployment stage and supports a later transition to supervised deep learning models as more data become available.

Energy management systems operating under dynamic electricity pricing require fast and cost-optimal control strategies for flexible loads. Mixed-integer linear programming (MILP) can compute theoretically optimal control trajectories but is computationally expensive and typically relies on accurate load forecasts, limiting its practical real-time applicability. This paper proposes a supervised imitation learning (IL) framework that learns optimal setpoint trajectories for a conventional proportional (P) controller directly from electricity price signals and temporal features, thereby eliminating the need for explicit load forecasting. The learned model predicts setpoint trajectories in an open-loop manner, while a lower-level P controller ensures stable closed-loop operation within a two-stage control architecture. The approach is validated in an industrial case study involving load shifting of a refrigeration system under dynamic electricity pricing and benchmarked against MILP optimization, reinforcement learning (RL), heuristic strategies, and various machine learning models. The MILP solution achieves a cost reduction of 21.07% and represents a theoretical upper bound under perfect information. The proposed Transformer model closely approximates this optimum, achieving 19.33% cost reduction while enabling real-time inference. Overall, the results demonstrate that the proposed supervised IL approach can achieve near-optimal control performance with substantially reduced computational effort for real-time energy management applications.

Energy management systems under dynamic electricity pricing require fast and cost-optimal control strategies for the optimization of flexible loads such as heating, ventilation, and air conditioning (HVAC) systems and refrigeration units. While Mixed-Integer Linear Programming (MILP) can compute theoretically optimal control trajectories, its practical application is limited due to computationally expensive optimization, leading to limited real-time applicability, and its dependence on accurate forecasts of electrical loads and other relevant time-series signals including disturbances. This paper proposes a supervised imitation learning (IL) framework that learns to imitate MILP-optimal setpoint trajectories for a conventional proportional (P) controller using only electricity price signals and temporal features. Our IL model predicts setpoint trajectories in an open-loop manner without direct state feedback and a subsequent conventional P-controller provides closed-loop robustness in a two-stage control structure. In this study, our approach is validated for electrical load shifting of a refrigeration system in an industrial warehouse, including a systematic benchmark of multiple IL models. MILP achieves a cost reduction of 21.07% relative to baseline and serves as a theoretical upper bound. Among IL models, sequence-based architectures achieve the highest savings, with Transformer and Long Short-Term Memory (LSTM) models closely approximating MILP behavior, reaching 19.33% and 19.28% respectively. A closed-loop reinforcement learning (RL) controller achieves 19.69% savings and is included as an additional benchmark, while heuristic strategies reach at most 14.43% savings. From a computational perspective, IL models enable fast training and real-time inference, with Transformer inference requiring 526 ns per prediction compared to 22.8 s for a single MILP optimization. This makes the proposed approach well suited for real-time and edge computing applications. Overall, the results demonstrate that the proposed supervised IL approach can achieve near-optimal control performance with substantially reduced computational effort, providing a scalable and cost-efficient solution for energy management.

As energy systems transition toward high shares of variable renewable generation, local energy communities (ECs) are increasingly relevant for enabling demand-side flexibility and self-sufficiency. This shift is particularly evident in the residential sector, where the deployment of photovoltaic (PV) systems is rapidly growing. While mixed-integer linear programming (MILP) remains the standard for operational optimization and demand response in such systems, its computational burden limits scalability and responsiveness under real-time or uncertain conditions. Reinforcement learning (RL), by contrast, offers a model-free, adaptive alternative. However, its application to real-world energy system operation remains limited. This study explores the application of a Deep Q-Network (DQN) to a real residential EC, which has received limited attention in prior work. The system comprises three single-family homes sharing a centralized heating system with a thermal energy storage (TES), a PV installation, and a grid connection. We compare the performance of MILP and RL controllers across economic and environmental metrics. Relative to a reference scenario without TES, MILP and RL reduce energy costs by 10.06% and 8.78%, respectively, and both approaches yield lower total energy consumption and CO₂-equivalent emissions. Notably, the trained RL agent achieves a near-optimal outcome while requiring only 22% of the MILP’s computation time. These results demonstrate that DQNs can offer a computationally efficient and practically viable alternative to MILP for real-time control in residential energy systems.

As energy systems transition toward high shares of variable renewable generation, local energy communities (ECs) are increasingly relevant for enabling demand-side flexibility and self-sufficiency. This shift is particularly evident in the residential sector, where the deployment of photovoltaic (PV) systems is rapidly growing. While mixed-integer linear programming (MILP) remains the standard for operational optimization and demand response in such systems, its computational burden limits scalability and responsiveness under real-time or uncertain conditions. Reinforcement learning (RL), by contrast, offers a model-free, adaptive alternative. However, its application to real-world energy system operation remains limited. This study explores the application of a Deep Q-Network to a real residential EC, which has received limited attention in prior work. The system comprises three single-family homes sharing a centralized heating system with a thermal energy storage (TES), a PV installation, and grid connection. We compare the performance of MILP and RL controllers across economic and environmental metrics. Relative to a reference scenario without TES, MILP and RL reduce energy costs by 10.06% and 8.75%, respectively, and both approaches yield lower total energy consumption and CO2-equivalent emissions. Notably, the trained RL agent achieves a near-optimal outcome while requiring only 22% of the MILP’s computation time. These results demonstrate that Double Deep Q-Learning can offer a computationally efficient and practically viable alternative to MILP for real-time control in residential energy systems.

High energy prices pose a significant challenge for the food industry due to their energy intensity. One promising method to reduce energy costs is demand response (DR). In this paper, we design a battery energy storage system (BESS) for DR in a food processing plant, which we use to shift the electrical load from high-price periods to low-price periods. A known challenge of load shifting is load forecasting. We compare extended long short-term memory (xLSTM), long short-term memory (LSTM), and Transformer with persistent prediction and historical average load profiles for load forecasting and evaluate the financial impact of forecasting errors in a case study. Despite using a small feature set, machine learning-based (ML) forecasts demonstrate strong performance, with symmetric mean absolute percentage error (sMAPE) values of 11.65% for xLSTM, 11.02% for LSTM, and 11.03% for Transformer, compared to 16.59% for historical average load profiles and 20.08% for persistent prediction. In the case study investigated, savings in electrical costs of 37,270 EUR per year could be achieved. This shows the enormous potential of BESS in combination with ML-based forecasting for industrial DR.