Maintenance Forecasting Model for Geographically Distributed Home Appliances Using Spatial-Temporal Networks

The application of machine learning in predicting regular and ad-hoc maintenance demand has been widely discussed recently. Reliable forecasting of uncontrollable, ad-hoc maintenance can improve resource allocation and spare part supply planning. However, the scope of its application is still limited to manufacturing and fleet management areas. Developing predictive analytics techniques for geographically distributed household appliances has been less discussed in literature. In this research, we propose a Spatial-Temporal Network (STN) model for forecasting ad-hoc maintenance needs of private home heating appliances considering external factors such as regional situation and weather. To evaluate the results of the model, we use five years of historical maintenance order data of a heating service company as training and test data. We compare the results of our model with traditional forecasting approaches like historical average, time series analysis and multi-factor linear regression. The evaluation results show a clear improvement of forecasting accuracy and outperformed the MAPE of the best traditional model by over 6%. The developed STN model provides the basis for implementing advanced prediction of maintenance requirements for uncontrollable, ad-hoc demand and offers a reliable demand planning base.