Forecasting Production Performance Using MLP and LSTM: Analysis of Design Approaches

This article examines the problem of production indicator forecasting using deep learning methods based on fully connected and recurrent neural networks. It focuses on the practical implementation of regression models by means of a multilayer perceptron (MLP) and a recurrent neural network with long short-term memory (LSTM), as well as on the analysis of their design approaches taking into account the characteristics of the source data. The work covers the full cycle of model construction: data collection and preprocessing, feature normalization, generating training and test samples, constructing neural network architectures, hyperparametertuning, training process, forecast qualityassessment, and resultvisualization. As an applied example, the problem of production volume forecasting based on real historical data and characterizing the performance of production facilities over a multi-year period is considered. For MLP models, the data is treated as independent observations, while for LSTM, fixed-length time sequences are generated, allowing for the dynamics and long-termindicator dependencies to be taken into account. A comparative analysis of architectural solutions, computational complexity, resource requirements, and forecasting accuracy is conducted. It is shown that LSTM demonstrates higher accuracy when working with time series due to its ability to account for temporal context, but requires significantly more computational effort. At the same time, MLP is characterized by its ease of implementation and faster training speed, making it suitable for tasks with poorly defined temporal data structure. The results of the study can be used in the design of intelligent information and analytical systems for the manufacturing sector and serve as a practical guide for choosing a neural network architecture with respect to the nature of the data and accuracy and resource requirements.