Forecasting energy efficiency in the fuel and energy complex based on the MLP model

Relevance. In the modern global economy, the demand for energy resources is steadily increasing, driven by the need to ensure energy security, support sustainable societal development, and simultaneously reduce the negative technogenic impact on the environment. At the state policy level, the main priorities are defined as the digitalization of industry – primarily the energy sector – the implementation of resource-efficient technologies, and the optimization of energy consumption. Under these conditions, accurate and reliable forecasting of energy efficiency becomes a key instrument for ensuring economic growth, increasing investment attractiveness, and supporting science-based strategic decision-making. Aim. To develop a methodology for forecasting energy efficiency based on the energy intensity indicator. The proposed methodology includes preliminary statistical processing of the source data using the log-normal distribution, validation of adequacy through the Kolmogorov–Smirnov criterion, and correlation analysis. At the next stage, hyperparameters are optimized, after which a Multilayer Perceptron (MLP) artificial neural network model is adapted and applied for evaluating and verifying the reliability of forecasts. The effectiveness of the methodology is demonstrated using information collected from 23 enterprises of the fuel and energy complex. Methods. Data processing was carried out using the log-normal distribution, while predictive validity was evaluated with the Kolmogorov–Smirnov test and correlation analysis. The number of neurons in the input, hidden, and output layers, as well as key hyperparameters (batch size, learning rate, activation function, solver, and number of epochs), were optimally selected based on the minimization of the MAPE error. To verify model stability, testing was performed on training and test datasets in a 20/5 proportion. Results. The developed MLP model demonstrated high forecasting accuracy: the error amounted to 4.35% on the training dataset and 1.21% on the test dataset. The model predicted a steady decrease in energy intensity up to 2030, when this indicator is expected to reach approximately 0.913 kg of oil equivalent per USD. These results confirm the potential for significant improvement in energy efficiency, cost reduction, resource savings, and enhancement of environmental sustainability. Conclusion. The method and model proposed in this study provide a reliable scientific basis for high-precision forecasting, strategic decision-making, and the implementation of innovative developments in the fuel and energy complex.