Variance‐Guided Evolutionary Reinforcement Learning for Multi‐Objective Energy Scheduling in Dairy Farms

Abstract

Electricity consumption on dairy farms has a direct impact on operational cost, renewable energy utilisation and load flexibility. Coordinating electric water heating and battery storage under dynamic tariffs, photovoltaic generation and temperature constraints forms a challenging multi‐objective sequential decision‐making problem. This paper proposes a variance‐guided evolutionary reinforcement learning (VG‐ERL) framework for joint water‐heater and battery scheduling. VG‐ERL combines PPO‐based policy learning, NSGA‐II parent selection, a shared‐critic multi‐actor architecture and variance‐guided safe mutation to preserve cost–thermal trade‐offs under a fixed interaction budget. The framework is evaluated in an Irish dairy‐farm setting and in a German case study using real electrical consumption data from dairy farms. On the Irish farm, VG‐ERL achieves the lowest net electricity cost and grid import, with zero temperature‐constraint violations and the highest Pareto‐front hypervolume among MORL methods. It reduces cost by 6% over rolling‐horizon MPC and 15.2% over rule‐based control. On the German farm, VG‐ERL maintains 99.6% thermal safety while reducing cost by 4.8% over rolling‐horizon MPC and by up to 21.4% over MORL baselines. Across both datasets, VG‐ERL provides a strong balance between cost reduction, thermal safety and Pareto‐front quality, making it a practical learning‐based approach for dairy‐farm energy scheduling.