Integrating PQTCN-MPC with innovation: A new strategy for accurate PV power prediction

Abstract

Accurately capturing long-term and short-term dependencies and cyclical relationships are core elements in determining photovoltaic prediction accuracy. In this paper, a novel method named PQTCN-MPC, which integrates a Parallel Quadratic Temporal Convolutional Network (PQTCN) with a Multi-Position Coding (MPC) Transformer, is proposed. First, PQTCN effectively extracts long and short-term depth dependencies of time series. Subsequently, the extracted features are encoded using MPC and Embedding, and the encoded features are concatenated. Finally, the output is obtained through an encoder and decoder structure. This study utilizes publicly available data from the Yulala solar system with three different resolutions. Ablation experiments validate that PQTCN-MPC enhances R² by 3.69 %, NRMSE by 21.11 %, and MAE by 48.35 % at a minimum. Experimental results indicated that PQTCN-MPC enhanced R² by 4.73 % under various seasonal conditions, while keeping NRMSE below 5 %, which underscores its high prediction accuracy and wide applicability.