Spatio-Temporal Interpolated Echo State Network for Meteorological Series Prediction.

Area of Science:

• Meteorology
• Data Science
• Machine Learning

Background:

• Spatio-temporal prediction in meteorology is crucial but challenging due to complex spatial and temporal interactions.
• Existing models are often computationally intensive, hindering practical application.
• Accurate forecasting of meteorological data like precipitation and air quality is vital.

Purpose of the Study:

• To develop an accurate and computationally efficient spatio-temporal prediction model for meteorological data.
• To address the challenges posed by the joint spatial and temporal effects in prediction.
• To provide an easy-to-implement alternative to complex existing models.

Main Methods:

• A novel spatio-temporal prediction model based on echo state networks (ESNs) was designed.
• Cubic spline interpolation was used to smooth and fill gaps in meteorological data.
• The elastic-net algorithm was employed for automatic selection and shrinkage of spatial variables within the ESNs.

Main Results:

• The proposed model achieved a normalized root-mean-square error of approximately 0.250 on both precipitation and air quality index datasets.
• The model demonstrated comparable accuracy to long short-term memory (LSTM) models.
• The developed ESN-based model exhibited significantly shorter computation times compared to LSTM.

Conclusions:

• The proposed echo state network model offers an effective and intuitive approach for spatio-temporal meteorological prediction.
• The model successfully integrates spatial and temporal effects, outperforming other methods in computational efficiency.
• This approach presents a promising neural network solution for predicting meteorological series accurately and efficiently.