Forecasting low voltage distribution network demand profiles using a pattern recognition based expert system

• Pattern recognition expert system to forecast demand profiles of an LV (low voltage) transformer.
• Incorporates ARIMAX forecasts, correlation clustering and NN (neural network) discrete classification.
• LV transformer load influenced by temperature, humidity and day of the week.
• Used ARIMA (autoregressive integrated moving average) modelling method with external variables to construct ARIMAX models.
• Variables included demand lags, moving average forecast, temperature and humidity.

The advent of distributed renewable energy supply sources and storage systems has placed a greater degree of focus on the operations of the LV (low voltage) electricity distribution network. However, LV networks are characterised by having much higher variability in time series demand meaning that modelling techniques solely relying on iterative forecasts to produce a next day demand profile forecast are insufficient. To cater for the complexity of LV network demand, a novel hybrid expert system comprised of three modules, namely, correlation clustering, discrete classification neural network, and a post-processing procedure was developed. The system operates by classifying a set of key variables associated with a future day and refining a recalled historical demand profile as the forecast. The expert system exhibited high hindcast accuracy when trained with a residential LV transformer's demand data with R2 values ranging from 0.86 to 0.87 and MAPE (mean absolute percentage error) ranging from 11% to 12% across the three phases of the network. Under simulated real world conditions the R2 statistic reduced slightly to 0.81–0.84 and the MAPE increased to 12.5–13.5%. Future work will involve integrating the developed expert system for forecasting next day demand in an LV network into a comprehensive distributed energy resource management algorithm.