Hybrid neural network-heuristic model for forecasting the energy demand of livestock farms

Abstract

Relevance. The article is devoted to the development and investigation of neural network and hybrid neural network–heuristic models for forecasting the energy demand of livestock farms. Purpose and objectives. The study aims to develop a hybrid neural network–heuristic model for forecasting the energy demand of livestock farms based on the integration of recurrent neural networks and heuristic optimization algorithms, enabling improved short-term forecasting accuracy under variable production and external conditions. The study develops a neural network model for forecasting the energy demand of a livestock farm based on recurrent architectures of long short-term memory and gated recurrent units, capable of processing multifactor hourly time series that characterize the operating modes of farm equipment. Methods. The baseline forecasting model is based on recurrent neural network architectures, including long short-term memory and gated recurrent units, while hyperparameter adjustment is performed using an evolutionary genetic algorithm. Model implementation and computational experiments were carried out using the Python programming language in a cloud computing environment. Results. The baseline neural network model demonstrated an acceptable level of forecasting accuracy, confirming its ability to reproduce the overall dynamics of energy consumption and seasonal variations. A hybrid neural network–heuristic model for forecasting the energy demand of livestock farms is proposed and formalized, combining a recurrent neural network forecasting module with a heuristic hyperparameter optimization algorithm. The use of an evolutionary genetic algorithm for automated tuning of the neural network architecture and training parameters is substantiated. This approach increased the model’s adaptability to changes in livestock farm operating regimes and reduced forecasting errors without a significant increase in computational complexity. As a result of the optimization procedure, an optimal neural network configuration was identified, characterized by an extended memory window, short-term forecasting horizon, and rational structure of hidden layers. The optimized model provided a noticeable reduction in forecasting errors compared to the baseline solution. To validate the effectiveness of the proposed hybrid model, computer experiments were conducted using data from a livestock farm. Conclusions. The obtained results confirmed stable forecasting performance both under average energy consumption regimes and peak load conditions, indicating the feasibility of applying the developed hybrid model as an analytical basis for decision support in the management of autonomous energy systems of livestock farms.