Investigation Neural Network Models for Wind Speed Prediction Based on Meteorological Observations in Northern Dagestan

Abstract

The paper presents the experiments' results on the study variants of neural network architectures for predicting wind speed based on meteorological time series for Northern Dagestan. When working with data and models, modern software tools of the Python programming language for Data Science are used, such as Keras — a library for modeling neural networks, Pandas — a library for processing tabular data, AutoKeras — a library for automatically generating a neural network by dataset, TabGan — a library for expansion a dataset with artificial data. As initial data, regular observations at the Kochubey (Northern Dagestan) meteorological station for the period 2011­2022 were taken with a frequency of generalization of measurements 8 times a day. The original semi-structured data is pre-processed and reduced to a structured CSV dataset format. The task of predicting wind speed is reduced to the task of classification, in which it is not the wind speed itself that is predicted, but the class number in accordance with the gradation. From the point of view of considering wind speed as a renewable energy resource, three classes with gradations are accepted: class 0: 0—3 m/s (quiet), class 1: 4—7 m/s (average wind sufficient for optimal wind turbine operating), class 2: 8 m/s and higher (strong wind). When performing experiments, the influence value on the prediction accuracy from several aspects were analyzed, such as: the time data block length, the neural network architecture, the transformation tabular features to normal form or to categorical form, expansion dataset by artificial data, the layout of the verification and test samples, the imbalance of classes, various meteorological parameters as features.