Research on Power Generation Forecasting for Photovoltaic Power Plants Based on Spatio-Temporal Multidimensional Feature Decomposition and Deep Learning Architecture

Abstract

Against the backdrop of energy transition driven by the “dual carbon” goals, this study addresses the challenges posed by the fluctuating power output of photovoltaic (PV) generation to grid stability by establishing an integrated forecasting framework combining physical modeling and deep learning. First, a theoretical irradiation and power generation model based on geographic location information was established as an ideal benchmark for deviation analysis. The STL decomposition method was applied to decompose the power sequence into seasonal, trend, and random components, while the K-means clustering algorithm quantified intraday power fluctuation characteristics. Analysis reveals a 55.30% deviation rate between actual and theoretical summer power, highlighting environmental uncertainty impacts. Subsequently, a “month-day-hour” three-dimensional tensor-based data reconstruction technique was introduced. By characterizing multi-scale features through seasonal indices and intraday distribution factors, the performance of LSTM and Transformer models in day-ahead power forecasting was comparatively validated. Experimental results demonstrate that the LSTM model exhibits significant advantages in capturing power variation trends, achieving a low root mean square error of 3.7229%. This effectively addresses the technical challenge of low prediction accuracy under complex meteorological conditions, providing reliable data support for grid dispatch.