Deep Deterministic Policy Gradient (DDPG)-based Photovoltaic Power Forecasting

The uncertainty and randomness of large-scale photovoltaic (PV) generation will seriously affect the safety and economy of the power systems. In order to mitigate the negative impact of grid-connected solar PV, it is important to accurately predict PV power. In view of this, this paper first proposes a novel hybrid method based on non-pooling convolutional neural network and deep deterministic policy gradient (DDPG) for ultra-short-term PV power forecasting. The DDPG model is used to implement the feature training task, and the NPCNN model is introduced into DDPG’s actor network to learn action strategies for partially observed problems in complex environments. The proposed method is extensively evaluated on real PV data in Limburg, Belgium. Numerical results show that the proposed method can provide good forecasting performance in different seasons and forecasting horizons.

I. Introduction

In recent years, photovoltaic (PV) power generation has exploded globally as a low-carbon renewable energy source due to global environmental degradation and depletion of traditional energy sources [1]. However, due to the irregularity and volatility of photovoltaic power generation, the grid connection of large-scale PV power generation has a significant negative impact on the stable operation and economic dispatch of the modern power system [2]. In this context, both PV power generators and grid dispatchers are required to accurately forecast PV power generation [3]. Therefore, it is an important task to use an advanced PV power generation forecasting method for optimal operation of the entire modern power system safely and economically.