Research and Development of Models for Intelligent Analysis of Big Data in Planning Tasks

The utilisation of machine learning techniques in the analysis of big data has become a standard practice for forecasting the outcomes of various sectors of the economy. This scientific article analyses changes in electricity consumption volumes in the Republic of Kazakhstan from 2002 to 2022. The study examines the dynamics of growth or reduction in consumption volumes. Data was prepared for analysis and model construction by employing outlier assessment methods. The work considers the Holt-Winters, SARIMA, and LSTM models, comparing their advantages and disadvantages. Accuracy metrics for forecasts and the Dickey-Fuller test for stationarity are also assessed. The application of forecasting models yielded diverse output data for predictive values, and it also examined the quality of training neural networks for analysing big data, specifically electricity consumption volumes. As an alternative, recurrent neural networks (RNNs) were utilised. The study revealed that this category of neural networks achieves results that closely approximate actual data and ensures precise training of neural networks for forecasting tasks. Moreover, the article explores the bidirectional RNN, a variation of recurrent networks that offers superior problem-solving quality in some instances. Recurrent networks process input sequences sequentially, and any alteration in data sequence can completely transform the representation extracted from the sequence. Hence, they excel in tasks where sequence order matters, such as temperature forecasting.

机器学习技术在大数据分析中的应用，已成为预测各经济领域发展结果的常规实践手段。本文针对2002年至2022年哈萨克斯坦共和国的电力消费量变化展开分析，重点考察了消费量的增减动态。研究通过异常值评估方法完成数据预处理，以支撑后续分析与模型构建工作。本文考量了霍尔特-温特斯（Holt-Winters）、季节自回归积分滑动平均（SARIMA）以及长短期记忆网络（LSTM）三类模型，并对比了各自的优缺点。同时还评估了预测准确率指标，以及用于检验序列平稳性的迪基-富勒检验（Dickey-Fuller test）。各类预测模型的应用生成了多组预测值输出结果，同时还分析了面向大数据（尤其是电力消费量分析场景）的神经网络训练质量。作为替代方案，本研究采用了循环神经网络（Recurrent Neural Networks，RNNs）。研究结果表明，该类神经网络可生成与实际数据高度贴合的结果，并能实现精准的预测任务神经网络训练。此外，本文还探讨了双向循环神经网络（Bidirectional Recurrent Neural Network）——循环网络的一个变体，其在部分场景下具备更优异的问题求解性能。循环神经网络按序列顺序处理输入数据，数据序列的任何调整都将彻底改变从序列中提取的表征信息，因此其在序列顺序至关重要的任务中表现突出，例如气温预测任务。