Modeling Turkey’s Hourly Electricity Market Clearing Prices Using Exponential Gaussian Process Regression

In the electricity market, the Market Clearing Price (MCP) holds strategic importance for both producers and consumers, as it represents the price formed at the equilibrium point of supply and demand. Accurate prediction of the MCP is critical for market participants in terms of production planning, cost estimation, and risk management processes. In this study, modeling was performed using four different machine learning (ML) methods, namely Exponential Gaussian Process Regression (EGPR), Rational Quadratic Gaussian Process Regression (RGPR), Quadratic Support Vector Machine (QSVM) and Medium Gaussian Support Vector Machine (MGSVM), in order to estimate the MCP in the Turkish Day Ahead Market (DAM). A total of 4001 data points were used for the training phase. Two methods were used to prove the robustness and generalizability of the models. These are: 10-fold cross-validation method, and testing with a data set (25 points) completely independent of the training data. The data used in the modeling process include real-time electricity consumption (MWh), hourly electricity generation by source (natural gas, hydroelectric power plants (DAM), lignite, run-of-river, solar energy), Purchase/Sale Bid Volume, Matched Purchase/Sale Quantity and U.S. Dollar Exchange Rate. The performances of machine learning methods were compared according to performance criteria such as Root Mean Square Error (RMSE), Mean Square Error (MSE), Mean Absolute Error (MAE) and Coefficient of Determination (R²). It was observed that the model created with EGPR had the highest R2 value of “0.908” and “0.913” for 10-fold cross-validation and an independent test dataset, respectively. In addition, it was obtained that the MSE, RMSE and MAE error metrics for the 10-fold cross-validation of the model created with EGPR were the lowest with 26026.20, 161.33 and 120.33, respectively. For the independent test dataset, the MSE, RMSE, and MAE error metrics were 7734, 87.94, and 61.56, respectively, resulting in lower prediction errors than other ML methods. It was observed from both the 10-fold cross-validation results and the independent test dataset results that the model created with EGPR made more successful predictions than the models created with RGPR, QSVM and MGSVM.