Towards Accurate and Efficient Power Load Forecasting: A Cooperative Game-Based Optimization

Power load forecasting has been widely used in the electrical industry, but its forecasting accuracy and efficiency tuning is challenging. Recent studies have attempted to address this problem using machine learning (ML). Nevertheless, there are two drawbacks. First, the prediction accuracy is heavily dependent on model hyperparameters. Second, model training on a large volume of data takes a long time overhead. To address these problems, we propose a novel approach called CogBO based on Bayesian Optimisation (BO) with cooperative games for the selection of the search space for hyperparameter weight analysis and Bayesian Optimisation during model training. CogBO uses two strategies. The first one is named WGS, which collects hyperparameters prior to model training and pre-analyses their impact on prediction accuracy through a combinatorial game. The second is called BOK, which is based on a Bayesian optimisation algorithm to repartition the parameter space by combining the degree of influence of the weights of the hyperparameters in WGS. We use a real test dataset to set up baseline experiments comparing linear analysis, regression analysis, and combinatorial game analysis with and without Bayesian optimisation for evaluating CogBO. The results in our experimental setting show that CogBO improves the prediction accuracy by almost 13% by the combinatorial game weighting of hyperparameters, while at the same time reduces time overhead by almost 30% on average,