To achieve successful business operations and optimize business processes, it is necessary to incorporate the request for forecasting as a key component in electricity markets. In the paper, based on the model developed at the Institute for Modeling in Energy Engineering of the National Academy of Sciences, the utilization of machine learning methods to predict the generation and consumption of electricity in Ukraine is proposed. The experiments were carried out involving the use of regression techniques to reconstruct a dataset from bilateral contracts to predict electricity generation and consumption. The results of employing machine learning algorithms on hourly peacetime data revealed the feasibility of accurately forecasting market volumes and thus electricity consumption one hour ahead, extending the predictive capability beyond a one-day planning horizon.
It is important to understand the market situation and the structure of supply and demand [
The hourly demand for DAM is the hourly amount of electricity that customers declare and want
to buy from DAM at the respective hourly prices. This amount directly affects both the marginal price
that will be created and the amount of electricity that will be bought and sold. According to the law of
supply and demand, the lower the demand for a product, the lower the price; the higher the demand,
the higher the price will be. This directly affects the volumes that will be accepted for sale for sellers
on DAM. There is a need to forecast the demand for DAM. But now in Ukraine, there are no
mechanisms, algorithms, or mathematical models that would allow forecasting with a low forecasting
error. The demand for DAM depends on a large number of factors [
To obtain forecast results as close as possible to real data, the scientists of the G. E. Pukhov
Institute for Modeling in Energy Engineering proposed the following [
1. Using trading results for the last available day; 2. Using the change in demand between the day on which the actual results are available and the day on which the forecast is made;
3. Taking into account the change in demand, for example, for the previous four weeks between the same days of the week.
A specialized algorithm for short-term hourly-based forecasting of the amount of electrical energy
that will be traded on DAM was proposed in [
In previous works [
Machine learning [
For its work, Scikit-learn uses the following popular libraries:
NumPy [
Scikit-learn specializes in machine learning algorithms for solving supervised learning problems:
classification (predicting a category with a limited set of valid values) and regression (predicting a
feature with real values) [
Regression algorithms [
The Scikit-learn library implements the following main methods: linear: models whose task is to construct a separating (for classification) or approximating (for regression) hyperplane;
metric: models that calculate the distance using one of the metrics between sample objects, and make decisions depending on this distance (k-nearest neighbors);
decision trees: training models based on a set of conditions optimally selected to solve a problem;
ensemble methods: methods based on decision trees that combine the power of many trees, and thus improve their performance, and also allow for feature selection (boosting, bagging, random forest, majority voting); neural networks: a comprehensive nonlinear method for regression and classification problems; SVM: a nonlinear method that learns to identify decision boundaries; Naive Bayes: forward probabilistic modeling for classification problems; PCA: linear method for dimensionality reduction and feature selection; t-SNE: a nonlinear dimensionality reduction method; k-means: the most common method for clustering, requiring as input the number of clusters into which the data should be distributed;
cross-validation: a method in which the entire dataset is used for training (as opposed to splitting into train/test samples), but training occurs multiple times, and different parts of the dataset act as a validation sample at each step. The final result is the averaging of the results obtained; grid search: a method for finding the optimal hyperparameters of a model by constructing a grid of hyperparameter values and sequentially training models with all possible combinations of hyperparameters from the grid.
In particular, the k-nearest neighbors algorithm [
Currently, experiments have been conducted on the application of machine learning methods for
forecasting the generation and consumption of electricity using the Scikit-learn open-source
software [
The volume of each of the four markets mentioned in the introduction is calculated for each hour of the day in megawatts per hour. The example shown in Figure 1 aggregates all data on electricity markets in Ukraine. For comparison, in Europe, market volumes are recorded every 15 minutes.
We can highlight the following tasks for which it is appropriate to use machine learning methods: reproduction of missing values in the database; hourly forecasting of the market for 24 hours in advance using previous data for several hours and the sliding window method for their sampling;
market forecasting with additional use of data for the previous year. 3.1.
Data from July 2020 to December 2021 were used to form the training and test datasets. The k-nearest neighbors method was used to populate the Bilateral column. Optimal hyperparameters are the following: the number of neighbors is 14, the metric is Manhattan, and the algorithm takes into account the distance to each neighbor. In the case of the four input parameters mentioned above, it was experimentally proven that reducing this set worsens the results.
Finding the optimal set of input parameters. The next step is to try using more parameters, in particular from the Bilateral column. The optimal number of input parameters for a machine learning algorithm is always limited. The fact is that when the number of input parameters increases, the algorithm begins to correlate random fluctuations in the input data with the output value. This effect is usually called the curse of dimensionality.
The set of input parameters obtained empirically is shown in Figure 2 and highlighted in black. The values Bilateral1 and Bilateral2 were used at the input of the machine learning model more than once. The optimal value of the hyperparameter in relation to the optimal number of neighbors has changed, now the value is 4. The results of data prediction for various algorithms are shown in Table 2. Figure 3 shows the deviations of the predicted values from the real ones for the test dataset. As can be seen from the graph, the probability distribution of the forecasting error is close to normal.
The intermediate prediction results when increasing the number of cells in the Bilateral column are shown in Figure 4. This research proves that it is optimal to use about 10–11 values from the available historical data for forecasting the bilateral contract market.
Data from July 2020 to December 2021 were used to form the training and test datasets. The k-nearest neighbors method was used to populate the Bilateral column. Optimal hyperparameters are the following: the number of neighbors is 14, the metric is Manhattan, and the algorithm takes into account the distance to each neighbor. In the case of the four input parameters mentioned above, it was experimentally proven that reducing this set worsens the results.
Finding the optimal set of input parameters. The next step is to try using more parameters, in particular from the Bilateral column. The optimal number of input parameters for a machine learning algorithm is always limited. The fact is that when the number of input parameters increases, the algorithm begins to correlate random fluctuations in the input data with the output value. This effect is usually called the curse of dimensionality. The set of input parameters obtained empirically is shown in Figure 2 and highlighted in black. The values Bilateral1 and Bilateral2 were used at the input of the machine learning model more than once. The optimal value of the hyperparameter in relation to the optimal number of neighbors has changed, now the value is 4.
The results of data prediction for various algorithms are shown in Table 2.
Figure 3 shows the deviations of the predicted values from the real ones for the test dataset. As can be seen from the graph, the probability distribution of the forecasting error is close to normal.
The intermediate prediction results when increasing the number of cells in the Bilateral column are shown in Figure 4. This research proves that it is optimal to use about 10–11 values from the available historical data for forecasting the bilateral contract market. 3.2.
The previous section analyses the prediction of bilateral market volume just one hour back. With an available and trained machine learning model, the iterative algorithm was used to restore the bilateral column for one year range. The results are shown below in Figure 5. The two lines displayed correspond to configurations of input parameters considered earlier.
The chart allows us to conclude that the process is globally sustainable, but eventually, it falls into loops and attractors. Unfortunately, these restored data cannot be reused to train subsequent machine learning models. The left side of the chart is quite different from the real data shown in Figure 6.
Several issues should be clarified within the framework of the study. First, what is the optimal depth of historical data (in hours) to use for forecasting. This depth may vary for different electricity markets. Another dilemma concerns the iterative approach. It is possible to forecast data 24 times one hour ahead, but it is also possible to forecast for all 24 hours at once, with a fixed state of the input window. 3.4.
Using data from the previous year in forecasting can be useful from several points of view. For example, the capacity of solar, wind, and hydro-generating stations depends significantly on the season. The capabilities of solar stations depend even on the time of day. Burning coal for environmental reasons is possible only when other energy sources are already working at full capacity. Data for the previous year already take into account flows of electricity from one market to another. However, the usefulness of last year’s data must be determined experimentally. 2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 1 0.99 0.98 0.97 0.96 0.95 0.94
We would like to thank the scientists of the G. E. Pukhov Institute for Modeling in Energy Engineering of the National Academy of Sciences of Ukraine for providing the data on electricity markets in Ukraine.
Effective business operations and the optimization of business processes necessitate the implementation of demand forecasting in electricity markets as a crucial component. We proposed the application of machine learning techniques for forecasting electricity generation and consumption in Ukraine. The experiments were carried out to apply regression methods in reconstructing a column using data from bilateral contracts, specifically to predict electricity generation and consumption. Results from employing machine learning algorithms on hourly peacetime data indicate the potential to forecast market volumes and thus electricity consumption one hour ahead with approximately 1% accuracy. The level of accuracy achieved should enable planning beyond a single day in the future. Appendix A. Fragment of the machine learning algorithm output_names = ['Bilateral'] feature_names = ['Bilateral11', 'Bilateral10', 'Bilateral9', 'Bilateral8', 'Bilateral7', 'Bilateral6', 'Bilateral5', 'Bilateral4', 'Bilateral3', 'Bilateral2', 'Bilateral2', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'Bilateral1', 'DayAhead1', 'Intraday1', 'DayAhead', 'Intraday'] print(F"Reading the datasets for bilateral column forecasting") bilateral_dataset = pandas.read_csv('datasets/bilateral-available-dataset.csv') bilateral_training = pandas.read_csv('datasets/bilateral-available-training.csv') bilateral_testing = pandas.read_csv('datasets/bilateral-available-testing.csv') training_inputs = bilateral_training[feature_names] training_outputs = bilateral_training[output_names] testing_inputs = bilateral_testing[feature_names] testing_outputs = bilateral_testing[output_names] print("Training the evaluation nearest neighbors regression model") evaluation_regressor = KNeighborsRegressor(n_neighbors = 4, weights = 'distance', algorithm = 'auto', p = 1, metric = 'minkowski', n_jobs = 8) evaluation_regressor.fit(training_inputs, training_outputs) predicted_training_outputs = evaluation_regressor.predict(training_inputs) predicted_testing_outputs = evaluation_regressor.predict(testing_inputs) print_evaluation_metrics(training_outputs, testing_outputs, \
predicted_training_outputs, predicted_testing_outputs) Appendix B. Fragment of the sliding window algorithm print("Populating the missing bilateral market data") forecast_source_path = 'datasets/bilateral-forecast-source.csv' print(F"Reading the dataset from file {forecast_source_path}") forecast_data_frame = pandas.read_csv(forecast_source_path) forecast_data_frame_length = forecast_data_frame.shape[0] for i in reversed(range(forecast_data_frame_length)): if (i % 100 == 0): print(i) if (i == forecast_data_frame_length - 1): bilateral11 = forecast_data_frame.iloc[i]["Bilateral11"] bilateral10 = forecast_data_frame.iloc[i]["Bilateral10"] bilateral9 = forecast_data_frame.iloc[i]["Bilateral9"] bilateral8 = forecast_data_frame.iloc[i]["Bilateral8"] bilateral7 = forecast_data_frame.iloc[i]["Bilateral7"] bilateral6 = forecast_data_frame.iloc[i]["Bilateral6"] bilateral5 = forecast_data_frame.iloc[i]["Bilateral5"] bilateral4 = forecast_data_frame.iloc[i]["Bilateral4"] bilateral3 = forecast_data_frame.iloc[i]["Bilateral3"] bilateral2 = forecast_data_frame.iloc[i]["Bilateral2"] bilateral1 = forecast_data_frame.iloc[i]["Bilateral1"] else: bilateral11 = bilateral10 bilateral10 = bilateral9 bilateral9 = bilateral8 bilateral8 = bilateral7 bilateral7 = bilateral6 bilateral6 = bilateral5 bilateral5 = bilateral4 forecast_data_frame.loc[i, "Bilateral11"] = bilateral11 forecast_data_frame.loc[i, "Bilateral10"] = bilateral10 forecast_data_frame.loc[i, "Bilateral9"] = bilateral9 forecast_data_frame.loc[i, "Bilateral8"] = bilateral8 forecast_data_frame.loc[i, "Bilateral7"] = bilateral7 forecast_data_frame.loc[i, "Bilateral6"] = bilateral6 forecast_data_frame.loc[i, "Bilateral5"] = bilateral5 forecast_data_frame.loc[i, "Bilateral4"] = bilateral4 forecast_data_frame.loc[i, "Bilateral3"] = bilateral3 forecast_data_frame.loc[i, "Bilateral2"] = bilateral2 forecast_data_frame.loc[i, "Bilateral1"] = bilateral1 day_ahead1 = forecast_data_frame.iloc[i]["DayAhead1"] intraday1 = forecast_data_frame.iloc[i]["Intraday1"] day_ahead = forecast_data_frame.iloc[i]["DayAhead"] intraday = forecast_data_frame.iloc[i]["Intraday"] forecast_data_frame.loc[i, "Bilateral"] = bilateral forecast_target_path = 'datasets/bilateral-forecast-target.csv' print(F"Writing the dataset to file {forecast_target_path}") forecast_data_frame.to_csv(forecast_target_path, index = False)