Search

The increasing integration of renewable energy, particularly solar photovoltaic (PV) power, presents challenges for power system operation. Accurate forecasts of renewable energy are both financially beneficial for electricity suppliers and necessary for grid operators to optimize operation and avoid grid imbalances. This paper proposes a forecasting framework to implement conformal prediction (CP) on top of point prediction models, which predict the PV power on a day-ahead basis, to quantify the uncertainty of those predictions. Simple and multiple linear regression, along with random forest regression, are used to construct the point predictions based on weather forecasts. Several variants of CP, including weighted CP, CP with k-nearest neighbors (KNN), CP with Mondrian binning, and conformal predictive systems, are built to transform the point predictions into rigorous uncertainty intervals or cumulative distribution functions to enhance reliability. The framework’s performance is evaluated using large datasets of weather predictions and PV power output in the Netherlands. Results indicate that CP combined with KNN and/or Mondrian binning after a linear regressor outperforms the corresponding linear quantile regressor. CP with KNN and Mondrian binning after using random forest regression demonstrates the most accurate probabilistic PV power forecasts, improving the weighted interval score by 14% compared to multiple linear quantile regression.

To be able to schedule the charging demand of an electric vehicle fleet using smart charging, insight is required into different charging session characteristics of the considered fleet, including the number of charging sessions, their charging demand and arrival and departure times. The use of forecasting techniques can reduce the uncertainty about these charging session characteristics, but since these characteristics are interrelated, this is not straightforward. Remarkably, forecasting frameworks that cover all required characteristics to schedule the charging of an electric vehicle fleet are absent in scientific literature. To cover this gap, this study proposes a novel approach for forecasting the charging requirements of an electric vehicle fleet, which can be used as input to schedule their aggregated charging demand. In the first step of this approach, the charging session characteristics of an electric vehicle fleet are translated to three parameter values that describe a virtual battery. Subsequently, optimal predictor variable and hyperparameter sets are determined. These serve as input for the last step, in which the virtual battery parameter values are forecasted. The approach has been tested on a real-world case study of public charging stations, considering a high number of predictor variables and different forecasting models (Multivariate Linear Regression, Random Forest, Artificial Neural Network and k-Nearest Neighbors). The results show that the different virtual battery parameters can be forecasted with high accuracy, reaching R2scores up to 0.98 when considering 400 charging stations. In addition, the results indicate that the forecasting performance of all considered models is somehow similar and that only a low number of predictor variables are required to adequately forecast aggregated electric vehicle charging characteristics.