Electricity Load Forecasting in Texas using Neural Networks to Enhance the Power Grid Stability

Long-term electricity load forecasting is important for achieving grid reliability and economic planning, especially in a state like Texas, in which the electricity consumption is sensitive to climate changes and the growing penetration of renewable energy sources. This thesis is concerned with enhancing the ability of the utility companies in Texas through an enhanced machine learning model to enable them to forecast energy demand and avoid a repeat of blackouts. Therefore, load forecasting is a key driver of energy distribution to ensure efficient control of peak load periods, and providing assurance on the readiness of the power grid to supply loads in the future. After Data Analysis, I found the variable tmpf – temperature that has the closest relationship with the target variable MWh and has the strongest positive coefficient. Moreover, the findings of perusing energy consumption characteristics with regards weekdays and weekend’s confirmed that temperature variations have much larger effects on electricity consumption than habitual activities. The same analysis also indicated that the hourly consumption is highest within 4pm and 8pm implying the need to accurately predict energy usage during this period especially when the weather is extreme. Through experimentation with various model architectures, I found that the proposed model RNN-LSTM-GRU outperformed other models including, RNN, RNN-LSTM, and RNN-GRU in terms of prediction accuracy. This model used a look back window of 30 time steps and achieved the lowest Mean Absolute Error (MAE) of 2519.89. Multiple look back windows (5, 20, 30, 50) were tested, with 30 providing the optimal balance between accuracy and complexity. I collected the electricity load data from the ERCOT and the weather data from the Texas ASOS Network for the duration of 2016-2023. I divided the data into training (2016-2022) and testing (2023) datasets. The proposed model was trained over 100 epochs, using a batch size of 10, early stopping with a patience of 20 epochs with the training dataset and was tested with the testing dataset. The minimum validation MAE was at epoch 29 with a value loss of 11583227 and training MAE of 5377.15. Preliminary time series plots also showed that the model eventually fit the test data quite well, although some fluctuations were apparent in the early steps of training. Based on this model, it is evident that it is possible to make highly accurate long-term load forecast in Texas, mainly because temperatures can be very volatile and the weather unpredictable, which make accurate energy demand forecasts critical. Through the use of this model, the utility companies can make provision for periods that experience high demands, control the generation of energy, and guarantee supply at a given period, thus increasing both effectiveness of the companies and reliability of the grid.