Multi-Horizon Photovoltaic Power Forecasting with an HHO-Optimized TCN-BiLSTM Model

Abstract

Accurate photovoltaic (PV) power forecasting requires models that align with operational horizons and use only information available at forecast issuance. This study develops a dual-input temporal convolutional network-bidirectional long short-term memory (TCN-BiLSTM) model with hyperparameters selected by Harris Hawks Optimization (HHO) and evaluates it using 33,408 observations from a 20-kW-class PV plant at 15-min resolution. Rolling direct forecasts are generated for H = 1 (15 min), H = 4 (1 h), and H = 96 (24 h), with H = 4 as the primary operational task. Historical measurements enter the historical branch, while target-period numerical weather prediction (NWP) and calendar variables populate the future-known branch; preprocessing uses training-fitted scaling and cyclic encoding. At H = 4, HHO-TCN-BiLSTM achieves a mean RMSE of 1.3726 ± 0.0205 kW and R2 of 0.9471 ± 0.0016 among six deep architectures. In an independent equal-budget H = 4 optimizer comparison, HHO yields the lowest test RMSE (1.370 ± 0.022 kW); GWO and GA are 1.39% and 2.34% higher, respectively, with Holm-adjusted p-values of 0.087 and 0.034. Architecture and optimizer rankings change at H = 1 and H = 96, and high-ramp events raise H = 4 RMSE to 3.3683 kW. The one-hour task provides the strongest empirical support for the proposed pipeline, while the cross-horizon results demonstrate substantial horizon dependence in both architecture and optimizer performance.