Attention-Enhanced CNN-LSTM with Spatial Downscaling for Day-Ahead Photovoltaic Power Forecasting

Accurate day-ahead photovoltaic (PV) power forecasting is essential for secure operation and scheduling in power systems with high PV penetration, yet its performance is often constrained by the coarse spatial resolution of operational numerical weather prediction (NWP) products at the plant scale. To address this issue, this paper proposes an attention-enhanced CNN-LSTM forecasting framework integrated with a spatial downscaling strategy. First, seasonal and diurnal characteristics of PV generation are analyzed based on theoretical irradiance and historical power measurements. A CNN-LSTM network with a channel-wise attention mechanism is then employed to capture temporal dependencies, while a composite loss function is adopted to improve robustness. We fuse multi-source meteorological variables from NWP outputs with an attention-based module. We also introduce a multi-site XGBoost downscaling model. This model refines plant-level meteorological inputs. We evaluate the framework on multi-site PV data from representative seasons. The results show lower RMSE and higher correlation than the benchmark models. The gains are larger in medium power ranges. These findings suggest that spatially refined NWP inputs improve day-ahead PV forecasting. They also show that attention-enhanced deep learning makes the forecasts more reliable. Quantitatively, the downscaled meteorological variables consistently achieve lower normalized MAE and normalized RMSE than the raw NWP fields, with irradiance-related errors reduced by about 40% to 55%. For day-ahead PV forecasting, using downscaled NWP inputs reduces RMSE from 0.0328 to 0.0184 and MAE from 0.0194 to 0.0112, while increasing the Pearson correlation to 0.995 and the CR to 98.1%.