Machine Learning and Optimization of the Solar Power Management in Remote Communities

The volatility of renewable energy technologies (e.g. wind, solar) stresses the control systems used to stabilizeelectrical grids and ultimately limits the amount of renewables that can be deployed. This problem isparticularly acute in smaller micro-grids that are typically powered by diesel generation, but also occurs inareas with high levels of renewables, e.g. Northern Europe and Hawaii. Electrical storage, coupled with somedegree of demand-response (i.e. the ability to control the timing of certain types of energy usage), can assist inmitigating this variability. Two specific problems need to be addressed: 1) in order to maximize the utility ofstorage and demand response, a control system with intelligence is required (though prescience would bepreferred!): if one could very accurately predict the shape of future supply and demand then optimal generation,storage and load control-schedules could be produced that minimize power fluctuations. 2) eliminating heavydependence of the design, commissioning and ongoing operations of these control systems, on participation oftechnical experts, especially at the smaller-scale of micro-grid communities. The Dalhousie team has extensiveexperience in building predictive models from heterogeneous data of the kind required by this application.Using data currently available from JAZZ we will identify feature sets (months 1-2) and build models (months3-6) to predict energy demand for customers. We will develop models for customers with similarcharacteristics (these similar groups will be determined by clustering). Models will be based on previous userecords, and will be informed by weather, time, type of day, special events, etc. Addressing the secondcompany problem, it is important to determine their optimal sizes, locations and configurations. This willinvolve combining Machine Learning methods with optimization techniques, and we will pilot a study of thisintegration.