Hybrid algorithm for medium-term photovoltaic power prediction. Two-stage clustering-CNN-BiLSTM-Attention: Case study Lima (Peru)

The dimensioning of a solar plant requires historical information on seasonal meteorological variability where it will be installed. To ensure efficient photovoltaic (PV) energy production, accurate predictions that recognize variability are necessary. There are various hybrid prediction models without clear verification of the quality of clustering. Current PV prediction models have architectures with more accurate results but with a significant increase in computational time. A novel hybrid model is proposed that uses real historical data with a two-stage clustering process: physical-deterministic clustering for climatic seasons and clustering with the K-Medoids method to segment overcast and clear days, the results of which are incorporated into the CNN-BiLSTM-Attention prediction model trained on the residual over the cluster-mean baseline, allowing competitive results to be obtained over time horizons of 1, 2, 4, 7, 14, and 21 days. The clustering quality is supported by the joint optimization of three indices: Silhouette Score (SS) between 0.6698 and 0.8616, Calinski-Harabasz (CHI) up to 655.86, and Davies–Bouldin (DBI) as low as 0.4320, identifying 𝐾 ∗ = 2 as optimal across all four seasons. The proposed model attains 𝑅 2 between 0.8330 and 0.8453, RMSE between 0.1585 and 0.1627 kW, and MAE between 0.1156 and 0.1171 kW across the six horizons evaluated under stratified 7-daily split, 5-fold expanding-window cross-validation, 95% bootstrap confidence intervals, and Diebold–Mariano testing. During the 2023–2024 Coastal El Niño event, the model performs ＋2.6% better in 𝑅 2 than under normal conditions (0.8721 vs. 0.8498), demonstrating the robustness of two-stage clustering under anomalous climatic regimes.