Skip to Content
EnergiesEnergies
  • Article
  • Open Access

20 July 2026

21 Pages

Dynamic Firm Capacity: Firm Dispatch of Utility-Scale PV–BESS Plants via Deep Reinforcement Learning and Controllable Curtailment

,
,
,
and
1
College of Science and Engineering, James Cook University, 1/14-88 McGregor Rd, Cairns, QLD 4878, Australia
2
Department of Electrical and Electronic Engineering, Xi’an Jiaotong-Liverpool University, Suzhou 215123, China
3
School of Electrical and Automation Engineering, Nanjing Normal University, Nanjing 210023, China
4
College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082, China
This article belongs to the Section A: Sustainable Energy

Abstract

This paper quantifies PV-based power dispatch availability, termed “dynamic firm capacity” (DFC), to provide operational certainty at the power plant level. Unlike conventional PV forecasts with unavoidable errors, DFC is a conservative dispatch target designed to be reliably achievable by the PV–BESS system, enabling firm generation. Curtailment is treated as a controllable plant-level resource to improve dispatch reliability. Three methods for determining DFC are compared: a baseline using PV forecasts directly, a fixed-ratio downscaling of the PV forecast, and a deep reinforcement learning (DRL) approach based on the Proximal Policy Optimisation (PPO) algorithm. Using data from Rugby Run Solar Farm (RUGBYR1) in Queensland, Australia, PPO achieves higher accuracy with smaller battery capacity and lower curtailment than the other two methods, with the advantage widening as battery capacity and power limit grow. Against two additional tuned benchmarks (an empirical probability-of-exceedance quantile forecast and a state-of-charge-aware rule-based policy), PPO reaches comparable dispatch compliance at a fraction of the curtailed energy. The best-performing PPO agents demonstrate 100% DFC realisation and near-zero dispatch error at near-zero curtailment, with a BESS capacity equivalent to one hour of peak PV output and a 0.67 E-rate power limit. Results are reported across 100 independently trained agents per scenario using the September 2023–August 2024 period, and the method’s advantage is preserved on a chronologically later, fully unseen evaluation window (September 2024–May 2025). An open-source DRL training environment is released to support reproducible PV–BESS research.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.