RSAM: Robust Self-Attention Based Multi-Horizon Model for Solar Irradiance Forecasting

With the widespread adoption of renewable energy sources in the smart grid era, there is an utmost requirement to develop prediction models that can accurately forecast solar irradiance. The stochastic nature of solar irradiance considerably affects photo-voltaic (PV) power generation. Since weather conditions have a high impact on solar irradiance; therefore, we need weather-conscious forecasting models to boost predictive accuracy. Although Recurrent Neural Networks (RNNs) has shown considerable performance in time-series forecasting problems, its sequential nature prohibits parallelized computing. Recently, architectures based on self-attention mechanism have shown remarkable success in natural language programming (NLP), while being computationally superior. In this paper, we propose an RSAM (Robust Self-Attention Multi-horizon) forecasting architecture, which mainly works in two parts: First, multi-horizon forecasting of solar irradiance using multiple weather parameters; Second, prediction interval analysis for model robustness using quantile regression. A self-attention based Transformer model belonging to the family of deep learning models has been utilized for multi-variate solar time-series forecasting. Using the National Renewable Energy Laboratory (NREL) benchmark datasets of two different sites, we demonstrate that the proposed approach exhibit enhanced performance in comparison to RNN models in terms of RMSE, MAE, MBE, and Forecast skill at each forecasted interval.

Automated summary

This study introduces a novel solar irradiance forecasting model utilizing self-attention mechanism in combination with multiple weather parameters and quantile regression to improve predictive accuracy and model robustness.