As a result of the energy transition, renewable energy sources (RES) have become a significant part of gross electricity generation in Europe. The growth of volatile generation goes along with increasing uncertainty and thus becomes a major challenge for market participants such as TSO, DSO and direct marketers. Using RES infeed forecasts for the Day-Ahead Congestion Forecast ensures that the expected consumption is covered in advance. Furthermore, by predicting the infeed level from RES, forecasts are a key element in building the market value of RES with direct impact on the market value of the day-ahead forecast. In addition, the forecast has an indirect impact on revenues. With decreasing quality, the costs per MWh in the intraday trading and for the balancing energy increases.

As demonstrated, a reliable and uniform infeed forecast is essential for various market participants. Sufficiently accurate infeed forecasts are already being offered by a large number of providers, using different approaches, methods and information sources leading to distinct results. So far, there is no forecasting model that dominates the other models in all respects. Due to the different origins of the forecast errors of the individual forecasts, there is no complete correlation between the forecast errors. Thus, the idea of making the optimal use of individual forecasts is to combine them. This could cause positive and negative forecast errors that compensate each other thus improving the forecasting quality. Thereby, the weighting of individual providers has a decisive influence on the quality of the combined forecast and thus on safe network operations as well as on profits of energy traders and producers. [1] Although combining forecasts is already a well-known method for improving forecast accuracies with a wide range of approaches such as the Bayesian methods, it is still underdeveloped. [2]

This paper focuses on a development of a multi-level hybrid concept and a comparison of at least five different methods for an optimized combination forecast. This includes methods of descriptive statistics as well as heuristic methods. Through dynamic and optimal weighting and the combination of the provider time series, the deviation of the overall forecast to the actual infeed is minimized. The methods continuously take into account the fluctuations in the forecast quality of individual providers. Firstly, weights for 3 provider forecasts are determined and compared using simple methods like least square method (LSM) and particle swarm optimization (PSO). Secondly, various hybrid model combinations are used to investigate whether a combination of the results of the individual methods can lead to an optimization of the overall result. An exemplary hybrid model includes in the first layer the determination of weights by, for example, two methods, PSO and LSM. In layer two, the determined weights of these two methods are further optimized and combined by means of neural networks. The development of the hybrid model therefore involves a variation of the order of methods and links. Furthermore, the number and type of applied methods is changed. To compare the methods, the hybrid as well as the simple methods are implemented and validated using RMSE and the same database of three provider time series. Based on [1], the hybrid models are expected to perform better than the simple methods.

References

[1] T. Schröter, A. Richter, M. Wolter, „Development of Methods for an Optimized Infeed Forecast of Renewable Energies”, 2018 PMAPS, June 24-28, Boise, 2018.

[2] J. Nowotarski, B. Liu, R. Weron, T. Hong, “Improving short term load forecast accuracy via combining sister forecasts”, Energy, vol. 98, pp. 40-49, 2016.