Energy cost and emission optimization in food industry using deep reinforcement learning and Transformer-based load forecasting

Given the substantial energy demands of the food industry, this study introduces a data-driven demand response strategy for industrial refrigeration in the food-processing sector, targeting reductions in electricity costs and equivalent CO 2 emissions. To this end, the thermal flexibility of a warehouse is quantified using real measurement data, resulting in an estimated thermal capacity of 182 kWh/K. A temperature bandwidth of 5 K as flexibility corresponds to an equivalent electrical energy storage of 445 kWh. To enable predictive control, a Transformer-Encoder architecture is used for short-term thermal load forecasting. By pretraining the Transformer on historical electrical load data from the entire factory, the model captures broader operational patterns, improving the forecast accuracy despite a limited amount of thermal data. Based on these forecasts, a deep reinforcement learning agent is trained to optimize the set point of a proportional–integral controller within predefined temperature limits. The agent successfully balances conflicting objectives, achieving up to 15.50% cost savings and 9.80% emission reductions compared to a baseline scenario. A Pareto front illustrates the trade-offs between cost and emission minimization. The results demonstrate the effectiveness of combining deep reinforcement learning with Transformer-based load forecasting to simultaneously reduce energy costs and equivalent CO 2 emissions, offering a valuable contribution to the path toward a net-zero industry. • Identified 182 kWh/K of thermal capacity from real data as flexibility for load shifting. • Plant-wide load data improved cooling demand forecasting via application of transfer learning. • Deep reinforcement learning handled forecast errors without reducing demand response potential. • Achieved up to 15.50% cost savings and 9.80% emission reductions via multi-objective optimization.