Multi-horizon short-term load forecasting using hybrid of LSTM and modified split convolution.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Data Science

Background:

• Accurate short-term load forecasting (STLF) is vital for power system stability, planning, and demand response.
• Existing STLF methods face challenges due to data non-stationarity and complex meteorological dependencies.

Purpose of the Study:

• To propose a novel hybrid technique, LSTM-SC, for enhanced single-step and multi-step STLF.
• To evaluate the model's capability in extracting sequence-dependent and spatial features for improved forecasting accuracy.

Main Methods:

• A hybrid model combining Long Short-Term Memory (LSTM) and a modified Split-Convolution (SC) neural network (LSTM-SC) was developed.
• The model was trained and validated using the Pakistan National Grid load dataset (NTDC) and public datasets (AEP, ISO-NE).
• The impact of meteorological features, specifically temperature, on forecasting accuracy was analyzed.

Main Results:

• The LSTM-SC model demonstrated superior performance on the NTDC dataset, achieving low error rates (RMSE, MAE, MAPE) for both single-step and multi-step forecasting.
• The model exhibited strong generalization capabilities when tested on AEP and ISO-NE datasets.
• Investigating the effect of temperature confirmed its significant impact on load forecasting accuracy.

Conclusions:

• The proposed LSTM-SC hybrid technique offers a robust and accurate solution for STLF.
• The model's ability to handle complex data dependencies and its generalization performance make it suitable for real-world power system applications.
• Further research can explore incorporating more features and advanced hybrid architectures for even greater forecasting precision.