The purpose of this paper is to improve the accuracy of solving prediction tasks of the missing IoT data recovery. To achieve this, the authors have developed a new ensemble of neural network tools. It consists of two successive General Regression Neural Network (GRNN) networks and one neural-like structure of the Successive Geometric Transformation Model (SGTM). The principle of ensemble topology construction on two successively connected general regression neural networks, supplemented with an SGTM neural-like structure, is mathematically substantiated, which improves the accuracy of prediction results. The effectiveness of the method is based on the replacement of the summation of the results of the two GRNNs with a weighted summation, which improves the accuracy of the ensemble operation in general. A detailed algorithmic implementation of the ensemble method as well as a flowchart of its operation is presented. The parameters of the ensemble operation are determined by optimization using the brute-force method. Based on the developed ensemble method, the solution of the task of completing the partially missing values in the real monitoring dataset of the air environment collected by the IoT device is presented. By comparing the performance of the developed ensemble with the existing methods, the highest accuracy of its performance (by the parameters of Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE) accuracy) among the most similar in this class has been proved.
An effective solution of the prediction tasks requires the high accuracy of the result with minimal resource and time costs for the operation of the chosen algorithm. In cases when high accuracy of the received result has the first priority, it is expedient to use ensemble learning. This paper describes a prediction method using a new, stacking-based GRNN ensemble model. Each member of the developed ensemble processes its own dataset, where the vectors of the original set of data are randomly shifted relative to the current point. The authors chose SGTM neural-like structure as a meta-algorithm for the formation of the result of the ensemble. This choice is argued by the high accuracy and speed of its work. The results of a number of experimental studies on the optimal parameters selection of the developed ensemble are described. A comparison of the efficiency of its work with a number of known predictors was done. Prospects for further research are described.