Methods for preventing overfitting in microclimate forecasting tasks

Abstract

The paper addresses the problem of overfitting in neural network models used for forecasting microclimate parameters in industrial facilities. It is shown that in microclimate control systems overfitting leads not only to reduced forecasting accuracy, but also to unstable control actions, increased energy consumption, and accelerated wear of actuators. The main focus is on NNARX-type neural network models, which use historical values of input and output parameters and are sensitive to limited and uneven training data. Practical methods for preventing overfitting are analyzed, including Dropout, weight regularization, and training data variation. The applicability of Dropout in the hidden layer of NNARX without violating autoregressive relationships is substantiated. It is shown that the combined use of these methods makes it possible to improve forecast stability, ensure smoother control signals, and enhance the reliability of intelligent microclimate control systems under real industrial operating conditions.