A novel probabilistic gradient boosting model with multi-approach feature selection and iterative seasonal trend decomposition for short-term load forecasting

Abstract

Existing regression, tree-based and NN models either lacks probabilistic prediction, takes longer training time, have high computational requirements or sacrifice accuracy. This paper introduces a novel framework, (MAFS+ISTD+PGBM), specifically to overcome these limitations. First three challenges are addressed by integrating gradient boosting and quantile regression model. The key idea is to combine speed and scalability of gradient boosting with probabilistic capabilities of quantile regression, forming PGBM. However, the issue of mediocre accuracy still remained. To address this, two pre-processing techniques are introduced. MAFS utilizes statistical methods and knowledge-based analysis to identify the most relevant features, while ISTD extracts and eliminates trend and seasonality components, ensuring stationarity. After rigorous evaluations, (MAFS+ISTD+PGBM) emerges as the superior performer surpassing all existing models in terms of training time and accuracy with highest R2 score of 0.997 and low values across all error metrics. The proposed model took less than one-third of training time (∼15 min) compared to CNN-LSTM+attn., (∼48 min), the only model with comparable accuracy of proposed model. Thus, proposed approach shall be used to empower grid operators with highly accurate and cost-effective probabilistic forecasts which allows them to make informed decisions about system stability and optimize resource utilization, ensuring reliability and efficiency.