Abstract: A long-term photovoltaic (PV) power forecasting model coupling signal decomposition with deep learning is proposed to address the insufficient accuracy caused by multi-meteorological coupling and strong non-stationarity. First, the maximal information coefficient (MIC) is used to mine nonlinear associations between meteorological features and PV power. Second, the power series is adaptively decomposed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the components are reconstructed into high- and low-frequency parts based on sample entropy. Then, an attention-based bidirectional long short-term memory network (AttBiLSTM) models the high-frequency component, while XGBoost models the low-frequency component point by point; the summed reconstruction is further refined by an XGBoost residual corrector. Experiments on 500 consecutive days of 15-minute data from a PV station in Gansu show that the proposed model achieves a mean absolute error of 1.149 6 kW, a root mean square error of 1.919 2 kW and a determination coefficient of 0.997 5, reducing the two errors by 31.7% and 43.7% compared with the PCC-LSTM-XGBoost model.