Implement tuned reinforcement learning in selecting different machine learning models prediction for peak electricity consumption

Electricity is one of the most impactful energy resources to a country's economy. Forecasting and planning for electricity is the key to success in both industrial and social aspects, especially in developing countries like Thailand. Accuracy in powerforecasting is essential to support growth in electricity demand, provide adequate power planning, and prevent power system outages. To support rising energy demand, offer enough power planning, and avoid power system disruptions, accurate powerforecasting is crucial. However, the qualities of the data and the nature of the issue have a significant impact on how accurate the various forecasting techniques are. Since more and more methods have been implemented to address the need for higher accuracy prediction, a sufficient number of models are currently available and there are rooms for developing ensemble solutions. This research implemented a modern approach with Double Deep Q-network (Double DQN), a Reinforcement Learning framework, as a model selector to assign the forecasting model with the lowest prediction error to each data point. The monthly peak electricity consumptions from 2003 to 2017 in Thailand recorded by the Electricity Generating Authority of Thailand (EGAT) were used to train and validate the proposed model. In addition, the hyperparameter tuning method based on Fractional Factorial Design is employed to reduce the tuning time. As a result, the Double DQN framework has generated the expected outcomes that beats all that of the individual models in terms of accuracy measured by mean squared error (MSE) by leveraging the strength of each forecasting model.