Reliable Grid Forecasting: State Space Models for Safety-Critical Energy Systems

TL;DR

The paper addresses safety-critical grid load forecasting by adopting grid-specific risk metrics that capture asymmetries between under-prediction and over-prediction. It evaluates Mamba-based state-space models on CAISO data, showing that standard accuracy metrics fail to reflect operational risk, while weather integration and tail-focused training yield substantial tail-risk reductions. The authors introduce S-Mamba, PowerMamba, and Mamba-ProbTSF architectures, compare them to iTransformer and Chronos, and demonstrate that S-Mamba achieves the lowest tail Reserve_99.5% margin (14.12%), with weather-enabled variants reducing large errors by up to ~40% and improving resilience under extreme conditions. The work also discusses practical deployment implications, including computational efficiency and the need for covariate-driven, weather-aware forecasting to support reliable, low-cost grid operations.

Abstract

Accurate grid load forecasting is safety-critical: under-predictions risk supply shortfalls, while symmetric error metrics mask this operational asymmetry. We introduce a grid-specific evaluation framework--Asymmetric MAPE, Under-Prediction Rate, and Reserve Margin--that directly measures operational risk rather than statistical accuracy alone. Using this framework, we conduct a systematic evaluation of Mamba-based State Space Models for California grid forecasting on a weather-aligned CAISO TAC-area dataset spanning Nov 2023--Nov 2025 (84,498 hourly records across 5 transmission areas). Our analysis reveals that standard accuracy metrics are poor proxies for operational safety: models with identical MAPE can require vastly different reserve margins. We demonstrate that forecast errors are weakly but significantly associated with temperature (r = 0.16, p < 10^{-16}), motivating weather-aware modeling rather than loss function modification alone. The S-Mamba model achieves the lowest Reserve_{99.5}% margin (14.12%) compared to 16.66% for iTransformer, demonstrating superior forecast reliability under a 99.5th-percentile tail-risk reserve proxy.

Figures (5)

• Figure 1: Quantile Loss Training Evaluation. (a) Error distribution comparison showing quantile loss training (orange) produces nearly identical distribution to MSE baseline (gray). (b) Large error counts by threshold---quantile loss reduces $>$1000 MW errors by only 1.6%, demonstrating that loss function modification alone cannot address large forecast deviations.
• Figure 2: Temperature-Error Association. (a) Mean absolute forecast error increases with temperature (slope = 24.1 MW/°C, $r = 0.16$, $p < 10^{-16}$). Error bars show 99.5% CI of mean. (b) Probability of large errors (top decile) increases during temperature extremes.
• Figure 3: Weather Integration Architectures. Each base model is extended to incorporate thermal-lagged weather covariates, motivated by the error analysis showing temperature-error correlation. (a) S-Mamba concatenates load and weather at input. (b) PowerMamba sums weather into seasonal/trend streams. (c) Mamba-ProbTSF uses summation fusion preserving uncertainty estimates. (d) iTransformer tokenizes weather variables for cross-attention.
• Figure 4: Weather Integration Benefit. Signed percentage forecast error distributions for each California utility and aggregate across all utilities. Each panel compares MSE baseline vs weather-integrated models on matched evaluation windows (baseline windows are sliced to exactly match the weather window set by exact target matching), with whiskers indicating the 0.5th to 99.5th percentile range.
• Figure 5: Quantile Loss vs Weather Integration Effectiveness (CA ISO-TAC, 24h). (a) Large error counts by threshold computed on matched evaluation windows. (b) Percentage change from the matched MSE baseline.