A Physics-Aware and SHAP-Guided Adaptive Method for PV Power Forecasting
- Purely data-driven PV models often predict negative power at night or low-irradiance segments, violating physics and hurting dispatch trust. Asymmetric non-negative penalties were embedded in Autoformer, FEDformer, and Informer backbones with ablations; negative-power predictions dropped sharply and baseline MSE improved steadily on DKASC.
- Models may over-rely on GHI and a few drivers, causing overfitting that is hard to spot in black-box training. SHAP tracking of GHI contributions was coupled to TPE Bayesian optimization for physically plausible hyperranges, flagging contribution drift mid-training and improving diagnosability and cross-season stability.
- Public DKASC streams drift across seasons and operating regimes, making full retraining expensive online. A SimpleADWIN-triggered module freezes backbones and fine-tunes terminal projections—~92% faster than full retraining with smaller post-drift loss—while mean SOTA MSE improved about 17%, supporting an integrated physics–explanation–evolution framework.
Abstract
Deep learning models for photovoltaic (PV) power forecasting often face challenges such as black-box opacity, physical law violations, and performance decay caused by concept drift. This research develops an adaptive forecasting framework guided by physics and Shapley additive explanations (SHAP). The proposed framework integrates domain physics with data-driven evolution. We use state-of-the-art (SOTA) forecasting architectures, including Autoformer, FEDformer, and Informer, as the backbone. We embed an asymmetric non-negative penalty into the loss functions to eliminate irrational negative power predictions. For diagnostic transparency, a SHAP-based introspection module monitors the global horizontal irradiance (GHR) contribution ratio to identify feature overfitting. This feedback directs a Bayesian optimization loop via the tree-structured Parzen estimator (TPE) algorithm to align model weights with physical causality. To ensure robustness in non-stationary environments, a lightweight evolution module triggered by SimpleADWIN freezes the SOTA backbone and fine-tunes the terminal projection layer. Validations on the Desert Knowledge Australia Solar Centre (DKASC) dataset demonstrate that the framework rectifies causal biases effectively. The method yields a 17% mean squared error improvement for SOTA models and accelerates online adaptation by 92% compared to global retraining.