An Advanced CBR-Based Photovoltaic Power Forecasting Framework Incorporating Kriging Meteorological Interpolation

Accurate meteorological data at photovoltaic (PV) sites are essential for reliable PV forecasting, but direct observations are often unavailable because of the spatial mismatch between meteorological stations and PV sites. This study proposes a Case-Based Reasoning and Ordinary Kriging (CBR-OK) framework for site-specific meteorological interpolation and subsequent PV forecasting. Historical variogram characteristics and model-selection results were organized into a case database, and ANN-derived feature importance was used to weight features for KNN-based case retrieval. An SVM trained on KNN-derived labels was then used to classify the variogram model for new cases, after which the corresponding historical variogram parameters were retrieved for OK interpolation. CBR-OK achieved interpolation accuracy comparable to Direct-OK, although end-to-end evaluation showed additional computational overhead from the CBR inference process. Using the interpolated meteorological inputs, LSTM, Bi-LSTM, and GRU models were evaluated for 1-, 12-, and 24-h-ahead PV forecasting, with the most consistent benefits observed at the 1-h horizon. As an example of absolute forecasting performance, the CBR-OK-based LSTM achieved mean absolute errors (MAEs) of 25.3, 45.7, and 47.3 kW at the 1-, 12-, and 24-h horizons, respectively, across the six PV sites. These results demonstrate the applicability of CBR-OK as an adaptive variogram-configuration and meteorological-interpolation framework for short-term PV forecasting.