Short-term wind speed forecast based on artificial neural networks and the method of variational mode decomposition

In the current conditions, climate change and the increasing frequency of extreme weather phenomena make the task of wind speed forecasting particularly relevant. In addition, short-term forecasting of local wind speed is extremely important to ensure safe and efficient operation of wind power stations and airports. Classical forecasting methods based on physical models of atmospheric processes are often inferior in accuracy to machine learning methods. Machine learning methods are able to efficiently process large amounts of data, detecting complex nonlinear dependencies. However, one of the main problems remains the presence of “noise” in the input data. This “noise” caused by external factors such as measurement error, turbulence, changes in temperature, humidity and other meteorological parameters, reduces the accuracy of the constructed models and, as a consequence, negatively affects the forecasting results. To solve this problem, approaches combining machine learning with data preprocessing methods are used. One of the promising directions is the use of artificial neural networks (ANN) combined with input signal filtering. In this paper, a hybrid method that combines neural networks with the Variational Mode Decomposition (VMD) method has been developed to improve the accuracy of short-term local wind speed prediction. This method allows to decomposed the input signal into several components (variation modes), each of which represents a certain frequency range, thus reducing the influence of noise and increasing the accuracy of useful information extraction. The method of decomposing the input signal into variation modes is applied to the input dataset (hourly measured values of surface wind speed) before using the ANN model for wind speed prediction. The aim of the work is to develop and apply a hybrid method for short-term prediction of local wind speed with an advance of up to 24 hours, which uses ANN in combination with pre-filtering of the input signal by VMD. Using historical wind speeds measured by a stationary weather station for the previous 24 hours decomposed into modes using the VMD, it is necessary to predict the wind speed in 1, 3, 6, 12 and 24 hours using an ANN. The ANN architecture is a classical fully connected neural network consisting of three layers: input, hidden and output layers. The size of the input layer is 576 neurons (24 time steps per 24 modes). Each neuron takes a numerical value corresponding to the characteristics of the modes decomposed using the VMD method. The hidden layer of the neural network contains 64 neurons that use the ReLU (Rectified Linear Unit) activation function. The output layer represents a single numerical value — the predicted wind speed in 1, 3, 6, 12 or 24 hours. The application of the hybrid method has achieved a significant increase in forecasting accuracy. In particular, the Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) decreased by at least 90 % (to 0.013–0.101 m/s and 0.9 %–6.1 %, respectively) for all considered advance options. The obtained values of the MAE and MAPE metrics confirm the high accuracy of the developed method, since a MAPE of less than 10 % can be classified as excellent prediction. In addition, the hybrid method shows high robustness to changes in data structure, which makes it a versatile tool for dealing with different types of meteorological conditions. The evaluation of the hybrid method results showed that the use of the VMD combined with ANN not only improves the quality of wind speed prediction, but also opens new opportunities for predicting other meteorological parameters. For example, temperature and humidity time series can also be processed using this approach, which will provide a comprehensive solution to the problems of meteorological analysis. The developed hybrid method for short-term wind speed forecasting is a promising tool that can significantly improve the accuracy of forecasts. Its application is especially relevant in the conditions of growing demand for reliable forecasts necessary to ensure the safety and efficiency of various weather-dependent systems. Further work in this direction can be aimed at improving the architecture of neural networks used within the method, as well as optimizing the VMD parameters. This will further improve the accuracy and adaptability of the models, which will make them indispensable in a wide range of tasks related to the analysis and forecasting of meteorological data.