Congestion forecasting using a custom loss function

Abstract

Whereas in the past, the Distribution System Operator (DSO) almost never encountered congestion in their grids, nowadays, with the increase of connected renewable energy sources, this will become more prevalent. To forecast congestion on transformer stations, with the goal of mitigating it, the Dutch DSO Stedin uses machine learning models with standard loss functions for regression. These fall short in predicting congestion peaks, as loss functions like MSE are unaware of the prediction goal, which is minimizing the cost associated with the congestion. Without knowledge of this goal, a loss function will not put any extra importance on finding congestion peaks and is more likely to miss them.

In this thesis, the use of a a custom loss function is explored, which incorporates the different costs associated with congestion mitigation. This function aims to improve upon existing congestion forecasts by having a loss function in line with the congestion forecasts goal. For this research, the case of substation Middelharnis is used. This substation encounters congestion due to connected wind parks and distributed photovoltaics. The congestion encountered will be solved by grid expansion in 2024, but until then will have to be mitigated using the redispatching market GOPACS.

A custom function was generated using two cost components: a congestion fee, which needs to be paid if wind parks are disconnected and GOPACS costs, which are the price of buying flexible power on the redispatch market. The function takes into account two-time horizons: costs associated with buying GOPACS power day-ahead acting on a prediction of the load, and the resulting cost on the day itself if any remaining congestion was resolved via the congestion fee. The contribution of this function is that it depends on the difference between the costs of the prediction and the cost of the realization, instead of the cost of the difference between prediction and realization.

Four models were trained, each with a different loss function: MSE, Pinball, and Cost. The cost model is trained twice and for each model, a different value for GOPACS cost is used. The results show that the cost models outperform the Pinball and MSE-trained models, when the performance is evaluated on the final cost metric. However, if the GOPACS price is high, the cost model becomes conservative in predicting congestion peaks. It is not expected that this will be an issue in the future, as with more renewable energy, the energy price and subsequently the GOPACS price will be lower.