Documentation for all RL algorithms and forecasting models used in the research. Each card includes version, framework, performance metrics, and a link to associated experiment runs.
Proximal Policy Optimisation
v3.2Best-performing algorithm. Achieves 18.6% cost reduction vs BAU. Stable training with clipped objective.
Soft Actor-Critic
v2.1Second-best performer. Entropy regularisation provides good exploration. Slightly slower convergence than PPO.
Twin Delayed Deep Deterministic Policy Gradient
v1.3Deterministic policy with twin critics for reduced overestimation bias. Good stability on investment decisions.
Asynchronous Advantage Actor-Critic
v1.0Asynchronous parallel workers. Faster wall-clock training but noisier gradients than PPO.
Trust Region Policy Optimisation
v1.0Constrained policy update with KL divergence bound. Theoretically motivated but computationally expensive.
Deep Q-Network
v1.0Discretised action space (5 levels per dimension). Included as baseline. Lower performance due to discretisation loss.
Long Short-Term Memory Forecasting Model
v2.12-layer LSTM for multi-horizon electricity demand and price forecasting. Test R²=0.9799 for annual demand.