A new forecasting system for electrical loads in the middle euphrates region using neural networks

This study develops neural network models for electrical load forecasting in Babylon Governorate, Iraq, focusing on Feedforward Neural Networks (FNN) and Long Short-Term Memory (LSTM) networks. Load forecasting is essential for managing power supplies and systems because it can guide the optimal allocation of resources, planning the right time for maintenance and reducing energy costs among others. Load forecasting is of great importance due to the complexity of power systems and the variability in demand. Statistical methods such as ARIMA have been used in traditional forecasting techniques. Although they work well in some circumstances, they do not incorporate all non-linear structures between variables. To address this, neural networks, especially those based on deep learning, have proven useful tools by enhancing forecasting accuracy. This thesis aims to use "FNN and LSTM", a form of artificial neural networks to forecast electrical load demand. The data analyzed in this research includes load data for 33/11 kV power distribution stations in Babil Governorate, Iraq, from January 2021 to April 2024. Temperature information was also collected from reliable sources such as Accuweather, Wunderground, and Weather. Load data and weather conditions were integrated into the model. The main activity is to forecast the electrical load for 2024 and 2025 with special attention to the effects of retaining fluctuations in demand during distinct months. Some of the data used in this research include monthly electrical load demand records as well as temperature records during the same period. The load data consists of energy consumption of several feeding stations located throughout the governorate, and temperature data is also used to understand the difference in energy use with the help of weather conditions. The first process was data cleaning and relevant data were formatted to improve data quality. MinMaxScaler and standardization methods were used where the data were normalized to scale for input into the neural network model. Normalization was applied, specifically, to make the ranges of all features equal because it is mandatory for the convergence of neurons during training. Model Design and Training, two different artificial neural network models were applied: the feedforward neural network (FNN) and the long short-term memory (LSTM) network. The aim is to analyze data and predict outcomes based on the nature of the available datasets. The selection of the optimal model depends on several factors, including data structure, model accuracy, and training efficiency. The FNN and LSTM models were compared, and loss and validation plots were constructed by tracing their respective loss curves during training and validation. The model was constructed, and model summaries were created. The FNN consists of dense layers with dropout regularization to prevent overfitting, resulting in a relatively lightweight architecture (approximately 11,521 parameters). However, the LSTM includes specialized layers for sequential data (such as LSTM and embedding), making it more complex (approximately 79,375 parameters). This complexity enhances its suitability for temporal or textual data but requires greater computational resources than the FNN. To validate and avoid overtraining the model, the Early Stopping technique was used to monitor the performance and stop training when the performance improvement stopped. The data obtained was divided into training and test datasets, where 80% of the data was trained and 20% was tested. Special methods, including the Dropout layer, were introduced to enhance the generality so that the model does not depend too much on the patterns in the training database. Moreover, hyperparameter tuning was also performed to fine-tune the "model performance," including learning rate, batch size, and number of epochs. The prediction accuracy of the neural network model constructed in the present work was measured by Mean Square Error in the test data (MSE) Mean Absolute Error (MAE) and Mean Percent Absolute Error (MAPE). These metrics provide a good indicator to evaluate the model's ability to predict unseen data and identify areas that deserve further improvement. The first set of results indicated that the model captured the general behavior of the electrical load demand pattern but with varying degrees of accuracy in different months. For example, there were higher demands during the summer, especially in July and August, as people used more air conditioning in various residential facilities. Similarly, demand in the winter months of January and February was higher due to heating systems. However, as temperatures dropped in April, May, October and November, the modeled demand was observed to adjust or decline. The model was particularly sensitive to these seasonal differences, hence the need to incorporate correct weather data into subsequent versions. Our results show a large approximation to the true values, differences in some months due to climate change, and large differences between summer and winter temperatures; temperatures range from a maximum of around 50 degrees in summer to a minimum of 5 degrees in winter. This is a large difference. This resulted in the test data evaluation being an approximate MSE of 95,000, MAPE of 18% in the test data evaluation. This indicates that the model can highly accurately predict and is useful in capturing broader trends. There are certain months that it is expected to produce lower forecast errors, and more factors should be considered when developing the model. More complex neural network architectures should be included in future studies to obtain better results due to some advantages of processing long series. Including more variables across multiple sectors, such as economic indicators, energy prices and consumption patterns, would make the demand model more accurate. The techniques used to produce ensemble estimates are generally superior to individual models because they can reduce variance and bias and thus produce more reliable estimates. Although the model has been shown to perform well, work and development in this area always need to be developed to keep pace with the increase in electrical loads and human growth; the model can be improved by incorporating neural network structures such as recurrent neural network / long short-term memory (RNN/LSTM) that can handle time series such as electrical load demand. These structures could cope with long-term dependencies in the data, which is necessary for forecasting over a long period. Another recommendation is to try adding more variables, including standard of living, economic factors, energy costs and energy consumption rates by different residential, industrial and commercial sectors. The first set of results indicated that the model captured the general behavior of the electrical load demand pattern but with varying degrees of accuracy in different months. For example, there were higher demands during the summer, especially in July and August, as people used more air conditioning in various residential facilities. Similarly, demand in the winter months of January and February was higher due to heating systems. However, as temperatures dropped in April, May, October and November, the modeled demand was observed to adjust or decline. The model was particularly sensitive to these seasonal differences, hence the need to incorporate correct weather data into subsequent versions.