Day-Ahead Photovoltaic Forecasting Based on Residual Downscaling and Multi-Scenario Dynamic RBF-LSTM Fusion

Abstract

To improve performance in day-ahead photovoltaic forecasting under irradiance fluctuations, meteorological conditions transitions, bias existed between numerical weather prediction (NWP) and local data measurements (LDM), as well the limited ability of a single forecasting model on global patterns capture in complex and highly variable data, a residual-downscaled multi-scenario dynamic fusion (RD-MSDF) framework is proposed. Firstly, historical paired data of NWP and LDM are used to train LS-Boost models, so to learn the residual features among physically matched meteorological variables, further add which to the corresponding NWP baselines to generate site-adapted meteorological features. After standardization, principal component analysis (PCA) is applied to reduce the dimensionality of corrected meteorological variables and reduce feature redundancy. The retained principal components are further combined with historical power data, fluctuation-related features, and temporal descriptors, so to construct unified forecasting inputs. Secondly, to characterize scenario-dependent operating conditions, a hierarchical scenario representation is developed. Rule-based hard scenarios are constructed according to corrected irradiance levels, irradiance fluctuations, day/night states, and historical power-fluctuation characteristics, while a Gaussian mixture model is employed to obtain soft-scenario posterior membership probabilities. Thirdly, LSTM and RBF network are adopted as complementary base predictors to capture temporal dependence and local nonlinear variations, respectively. Their preliminary outputs are calibrated through a soft-scenario mixture-of-experts module and subsequently fused by a hard-scenario-based Scene-EWAF++ mechanism, which maintains independent fusion-weight buckets for different operating scenarios and updates the active weights according to recently observed prediction errors. Finally, the proposed framework is evaluated with a public dataset from a 20 MW photovoltaic plant across four seasonal test sets. The comparison results show that it can obtain lower forecasting errors than other methods under varying irradiance and power-fluctuation conditions. The observed error reduction can effectively save the backup capacity for scheduling, as well potential cost of supply and demand power balance dispatch and carbon emission.