An Hour-Specific Hybrid DNN-SVR Framework for National-Scale Short-Term Load Forecasting.

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Short-term load forecasting (STLF) is crucial for power system stability and efficiency.
• Accurate forecasting requires sophisticated models that can handle complex, non-linear patterns.
• Integrating diverse data sources, including weather predictions, is essential for improving forecast accuracy.

Purpose of the Study:

• To develop and evaluate a hybrid deep neural network (DNN) and support vector regression (SVR) architecture for national-scale, day-ahead STLF.
• To propose an hour-specific framework comprising 24 hybrid models for enhanced forecasting precision.
• To integrate global numerical weather prediction data with local measurements for a realistic operational setup.

Main Methods:

• A hybrid DNN-SVR architecture was developed, with DNNs learning nonlinear representations and SVR performing final regression.
• An hour-specific framework with 24 hybrid models was implemented.
• Gradient-boosting for feature selection and integration of TIGGE weather data with local load and meteorological data were employed.

Main Results:

• The proposed hybrid model consistently reduced forecasting errors compared to standalone DNN, LSTM, and Transformer baselines in the test year 2022.
• The framework demonstrated a reproducible pipeline for STLF.
• The model successfully integrated global weather forecasts into a practical day-ahead STLF solution.

Conclusions:

• The hybrid DNN-SVR framework offers a significant improvement in STLF accuracy for real-world power systems.
• The hour-specific approach and integration of global weather data enhance the robustness and reliability of forecasts.
• This study provides a systematic comparison and a practical solution for national-scale day-ahead STLF.