The concept of the solution is based on two hypotheses.
First, DHS plant efficiency can be improved with: 1) using accurate demand and supply forecasting models, developed by taking a holistic approach in modeling plant operation, distribution system and consumer behavior; and 2) implementing effective real-time monitoring, facilitating timely and adequate (with supporting explanation) response to non-desired behavior of the system.
Second, the heat demand can be optimized through assisted, inclusive and transparent operation of the DHS plant and associated facilities, as well as responsible behavior of the end consumers.
The proposed concept of the solution is illustrated on figure below.
The central model in the architecture is the AI demand/supply forecasting model. Model aims to represent the behavior of complete DHS system. It is trained with historical data about heat load and other relevant features collected in DHS plant, substations and at the individual consumer level (sample). Besides water temperatures in supply and return, primary and secondary lines, model features are weather data, room ambient temperatures and air qualities (at individual consumer level). Model is developed by training the selected, optimized architecture with the acquired data. The model forecasts heat load (demand) and temperature of water on primary and secondary supply lines (supply), based on weather data and desired ambient room temperature. In exploitation, forecasting models will be used for: 1) short-term (e.g. hourly) and long-term demand and supply forecasts in the DHS plant operation; 2) dynamic calculation of supply control parameters which will then be used for defining hot water supply set points – control points for SCADA air temperature offset function. In short-term, rolling forecast strategy is implemented, meaning that model is re-trained every hour with actual data acquired from the DHS plant. In the exploitation, each short-term forecast will be associated with a local interpretation in the form of different explanation families and used for revising plant operation strategy (increasing and decreasing supply line water temperature). Long-term forecasts will be used for supply chain planning, issuing alerts and what-if analyses.
For the effective control, recognizing expected or unexpected unwanted system behavior is of critical importance. This will be achieved by implementing real-time data pattern matching and anomaly detection strategies on data streams incoming from sensors. Example data patterns indicating issues are: 1) those, where temperature of secondary return line water is not invariant (or not with minimal variance) and lower or higher than 45oC, 2) those, where difference between temperatures of secondary supply and return line water is increasing. Those patterns will be detected by using sliding window approach for bulk historical data. While expected unwanted behavior data pattern detection is a supervised problem, anomaly detection aims to discover unexpected faulty behavior. Some of the examples of such behavior are leakage, clogging or blockage of heat exchangers or pipes, faulty insulation, sensor bias and/or drift, open or close failures of valves and actuators and others. One of the selected ML/DL techniques will be implemented for anomaly detection.
The effect of the detected anomalies and expected unwanted behavior (data patterns) can be quantified by comparing the data about actual and alternative operation strategies. The dataset representing alternative scenario will be created by removing the outlier data and imputing the new ones, by using the existing demand/supply forecasting model. Comparation of the calculated costs and environmental effects of actual and forecasted (alternative) strategies would take place then. Finally, such analysis can be supported with the explanations for the occurrence of unwanted system behavior and confidence levels . This approach will also be applied on a bulk historical data to highlight and explain issues, suggest and apply alternative strategies (what-if analyses) to quantify and compare operation cost and environmental effects of implemented and alternative operation strategies, justify investments in control equipment, etc.
The objective of the end user prediction model is to help optimizing the demand, namely, to facilitate responsible citizens’ behavior through the best-informed personal choices related to the effect of different factors to the room temperature. End user prediction model is regression (ML/DL) model for predicting room temperature based on the DHS heating parameters (supply line water temperature and flow in DHS substation), weather information, room configuration, seasonal factors (time of a day, day in a week, month) and air quality (indicating open windows).
All the functionality supported by the proposed solution will be implemented in XAI4HEAT Plant Operation software and XAI4HEAT mobile app to the citizens. Key functionalities of the XAI4HEAT Plant Operation software will be:
- demand/supply rolling and long-term (for a given horizon) forecasts with economic and environmental impact analysis, associated with explanations;
- alerts with detected anomalies and unwanted behavior and associated explanations – root-cause analyses;
- what-if analyses’ tools for planning DHS system operation strategies;
- historical data analysis of the economic and environmental impact of different alternative operation strategies.
Mobile app will facilitate:
- insight into current and forecasted DHS plant operation parameters with interpretations, including fuel consumption and environmental effects (waste, CO2 emissions);
- visualizations of the effect of different factors to the room temperature;
- what-if scenarios in which heat consumption and effect to DHS resource consumption and environment in different behavioral circumstances (open window ventilation, additional electrical heating, etc.) can be calculated;
- different alerts with recommendations for consumer demand optimization.