Day-ahead Grid Loss Forecasting : A study of linear and non-linear models when modelling electrical grid losses

Abstract: Accurate day-ahead grid loss forecasts are, among other things, essential to determine the electricity price for the upcoming day. The more accurate forecast, the closer the trading on the 'day-ahead' electricity market can become the actual operation the next day, which dedcrease the need for correcting production on the balancing market. Followingly, the need for extra imbalance costs, which make the electricity price higher, is reduced with accurate forecasts. This project's purpose was to explore a wide range of mathematical models to increase the energy comapny Fortum's day-ahead grid loss forecasting accuracy, and thereby contribute to lower the risk for high imbalance costs.  Two electrical grids located in Sweden, with different characteristics, were studied. One electrical grid was located in Dalarna and the other one was located in Värmland. Four different model types were tested for each grid. The linear models ARIMAX and SARIMAX were explored and the two artificial neural networks FNN (Feedforward Neural Network) and LSTM-RNN (Long-SHort Term Memory Recurrent Neural Network) were explored. By constructing different model structures of each model type, as well as statistically testing which predictors to include as input to the models, the most accurate model for grid loss forecasting was found. The models' forecasting accuracy were validated based on the MAPE (Mean Absolute Percentage Error). Variables important as predictors were found to be power production, electricity prices and grid losses at previous time steps. For the grid in Dalarna ARIMAX(2,0,2) was the model generating most accurate day-ahead grid loss forecasts and for the grid in Värmland, SARIMAX(1,0,0)(0,0,1)[24] was the most accurate model. That is, different models were found as the most accurate one for grid loss foracsting, as the two studied electricity grids had different characteristics. Hence, this result implies that there is no universal model that is the most adequate at modelling all types of grid losses. To find useful models when forecasting grid losses day-ahead, an analysis of the particular grid losses being studied is therefore not irrelevant.