Compute-Constrained Adaptive Electricity Load Forecasting via Offline Policy Learning

Background: Electricity load forecasting services must allocate inference across models with different accuracy and compute costs under a shared daily budget. Methods: We evaluate offline policies using only pre-call information, complete-day rolling folds, outcome-time embargoes, and a reservation rule that preserves minimum service for later origins. Four public datasets are tested at four horizons and three budgets, comparing contextual allocation with fitted Q iteration (FQI), double FQI, and neural FQI against a feasible static forecaster. Results: All 960 formal policy evaluations satisfy the budget. The mean gains in mean absolute error (MAE) are 9.22%, 7.37%, 7.77%, and 3.83% for the four methods, respectively. The contextual gains are positive on Steel, Tetouan, and Open Power System Data (OPSD) but negative on UCI ElectricityLoadDiagrams (UCI–ECL). A disjoint deployment gate yields a 10.63% aggregate gain at 51.7% coverage with six harmful activations and 61 missed beneficial cases; this empirical risk-control filter is not a safety guarantee. The lightweight multilayer perceptron (MLP) extension and sequential audits find no uniform benefit from a larger action pool or deeper Bellman backups; FQI improves over contextual allocation only at the smallest tested switching penalty in the cache-state experiment. Conclusions: Offline policy learning can support budget-feasible adaptive forecasting, but its benefit depends on dataset and runtime conditions. Chronological validation, explicit feasibility checks, and static fallback are needed; the results do not establish the universal superiority of sequential reinforcement learning.