Comparative analysis of short-term weather forecasting models for autonomous observation systems

Under conditions of limited computational resources, the task of short-term weather forecasting (temperature, pressure, wind speed, and direction) is addressed using various methods. This work compares three approaches—naive model, regression models (Holt’s and linear regression), and the LSTM neural network—based on data from an autonomous weather station, for 24- and 72-hour forecasts. Results indicate that exponential smoothing provides the best balance between accuracy and computational efficiency, while the LSTM neural network approach achieves the highest forecast accuracy when sufficient resources are available.