A Hybrid Fourier–Deep Learning Forecasting Model for Solar Photovoltaic Power Using an Adaptive Cleanliness Index

Abstract Accurate solar photovoltaic (PV) power forecasting is vital for efficient renewable energy integration. Existing models often overlook key factors such as module cleanliness and dynamic atmospheric conditions. This study introduces a hybrid atmospheric-interferometric model (HAIM) that integrates meteorological data, spectral analysis, and an Adaptive Cleanliness Index (ACI) to improve short-term PV power forecasting. The proposed framework incorporates a Fourier Analysis Network (FAN) and dual convolutional neural network–long short-term memory (CNN–LSTM) channels with feature fusion. The evaluation was conducted using a real-world PV dataset comprising 34 days of hourly data from two solar power plants in India, without seasonal coverage. Experimental results show that HAIM reduces mean absolute error (MAE) by 30.50% and root mean square error (RMSE) by 10.72% and improves the coefficient of determination ( R 2 ) by 20.29% compared with those of the baseline FAN model ( p < 0.001 ). The ACI module provides its strongest contribution at the 1-h forecasting horizon ( t + 1 ), improving MAE by 10.73% ( p = 0.0005 ), with particularly strong gains given low-temperature conditions. However, the effect of ACI is horizon- and condition-dependent rather than uniformly positive across all samples. Economic and environmental assessments indicate projected annual savings of EUR 8.93 million and a reduction of 25.01 million kg CO 2 . These findings suggest that HAIM provides promising short-term, site-specific forecasting performance for the studied PV plants, while further validation using longer-duration, multiseason, and multisite datasets is required before confirming broader operational generalizability.