Degradation-Aware Deep Reinforcement Learning for Battery Energy Storage Dispatch in Industrial Microgrids and Its Real-Time Deployment Validation

Battery degradation is a non-negligible cost in battery energy storage (BESS) dispatch for industrial microgrids, yet most deep reinforcement learning (DRL) studies optimize only the procurement cost. This paper embeds a rainflow-calibrated cyclic aging cost into the reward of a degradation-aware soft actor-critic (DA-SAC) framework with an adjustable weight λ, making the cost–lifetime trade-off explicit. On one year of 15 min data from an industrial microgrid in Chongqing, the main evaluation is conducted in an offline simulation, and the typical days are validated by hardware-in-the-loop (HIL) real-time simulation on the Speedgoat real-time target with the policy on a Beckhoff edge controller. DA-SAC reduces the cyclic degradation cost by about 28% relative to MPC, but its operating cost cannot go below CNY 649.4/day, worse than a simple price threshold rule (641.8), a boundary we attribute to long-horizon credit assignment. Dynamic programming (DP) provides a strong baseline whose closed-loop lookup reaches cost parity with MPC under noisy forecasts (CNY 618.4 vs. 618.1/day), and deployment time SAC correction improves the coarse-grid DP from CNY 620.6 to 619.4/day, with greater forecast noise robustness (+3.5% vs. +4.55% for MPC under a three-fold noise increase). Degradation-aware DRL is therefore most valuable when combined with the exact opportunity cost of DP under a single main site and a fixed forecast noise level.