Short-Term Load Forecasting for Electrical Power Distribution Systems Using Enhanced Deep Neural Networks

Abstract

The rationale for using enhanced Deep Neural Networks (DNNs) in the power distribution system for short-term load forecasting (STLF) originates from a thorough analysis of current trends, the emergence of the state-of-the-art use cases and approaches. STLF plays a crucial role in economic load dispatch, hydrothermal coordination, system security assessment, load shedding, unit commitment, and cost-effective risk management in power systems with renewable energy sources. In this study, we introduce a Long Short-Term Memory (LSTM), augmented with enhancements inspired by an Efficient and Parallel Genetic Algorithm (EPGA) for the STLF of Jimma town power distribution system. To forecast the load in the short term, the model takes into account wind direction, wind speed, humidity, temperature, season, load history, and peak load due to holidays. The optimal linear combination of inputs for determining daily load is derived using EPGA and data from Ethiopian Electric Utility (EEU). The linearly combined data is then fed into the LSTM model for load prediction. During training, this allows the LSTM model to focus on the pattern of a single time-series data rather than the best combination of many input patterns. The proposed method uses the combination of EPGA and LSTM models for accurate STLF. Thorough experimental analysis indicates that the root-mean-squared error (RMSE) achieved by the EPGA-enhanced LSTM for Jimma town power distribution system STLF is about 43.87. This represents a 7.486% improvement over the prediction obtained using only LSTM model. Additionally, the mean average percentage error (MAPE) for five sample loads used to test the EPGA-enhanced LSTM prediction method further supports the robustness of the proposed method.

Keywords

• Deep neural networks
• long short-term memory
• load forecast
• short-term load forecast
• efficient and parallel genetic algorithm
• EPGA enhanced LSTM