Evolving Large Scale Prediction Models for Vehicle Volume Forecasting in Service Stations

Resource Planning and Service Optimization for operational efficiency constitutes a major factor in the service industry. Internally most of it is dependent on the accuracy of the forecasted demand for the service, which is used to proactively plan resources to match expected demand. In this paper, our focus is on a real-world scenario of vehicle volume forecasting in service stations. Previous work has explored a genetic algorithm (GA) to evolve a regression model based on Neural Networks. Our focus here is to extend on this and show that GA based approach can be also used to evolve other popular regression models for this problem that are widely used in machine learning literature. Each of these techniques considers the historical vehicle volume data along with other correlated data, such as weather, and can have its own set of model parameters as well as other parameters related to data filtration, correction, and feature selections. All of these parameters require proper tuning to achieve the best forecasting accuracy. This can be a challenging task, particularly where different prediction models need to be built for different stations and for different periods, potentially resulting in hundreds of models being built. Manual tuning can be time-consuming, and most importantly, sub-optimal. Our results show that GA can be successfully used to automate the optimization of many popular machine learning models for large-scale vehicle volume forecasting, and more importantly can provide better accuracy than traditionally used manual tuning approaches.