Long Short-Term Memory-Based Twin Support Vector Regression for Probabilistic Load Forecasting.

Area of Science:

• Power Systems Engineering
• Computational Intelligence
• Data Science

Background:

• Accurate and reliable probabilistic load forecasting is essential for efficient power system operation and energy resource management.
• Estimating uncertainties in forecasting models and non-stationary electric load data presents a significant challenge.

Purpose of the Study:

• To propose a novel probabilistic load forecasting model, BFEEMD-LSTM-TWSVRSOA, designed to estimate uncertainties in forecasting models and non-stationary electric load data.
• To evaluate the performance of the proposed model against various machine learning and deep learning algorithms using the Global Energy Forecasting Competition 2014 dataset.

Main Methods:

• The proposed model integrates Fast Ensemble Empirical Mode Decomposition (FEEMD) for data filtering, Long Short-Term Memory (LSTM) networks for feature extraction, and Twin Support Vector Regression (TWSVR) for forecasting.
• Parameters are optimized using Seeker Optimization Algorithms (SOAs).
• Bootstrap methods and forecasting step sizes were analyzed to determine optimal prediction intervals and point forecasting results.

Main Results:

• Experimental results on four seasonal datasets from GEFCom2014 indicate that the wild bootstrap method and a 24-h step size are optimal for the proposed model.
• The BFEEMD-LSTM-TWSVRSOA model achieved significant improvements, outperforming suboptimal models by an average of 46% in point forecasting and 53% in probabilistic forecasting across the datasets.

Conclusions:

• The proposed BFEEMD-LSTM-TWSVRSOA model demonstrates superior performance in both point and probabilistic load forecasting compared to existing methods.
• The study validates the effectiveness of the FEEMD, LSTM, TWSVR, and SOA components in enhancing load forecasting accuracy and reliability.