DEVELOPMENT OF A REAL-TIME ELECTRICITY DEMAND FORECASTING PLATFORM BASED ON MACHINE LEARNING AND APACHE KAFKA STREAM PROCESSING

I. Khussainov, D. Astapenko, Gulnar Kim, A.A. Kabdirova

2026

Closed access Metadata (CSL-JSON)
DOI 10.51775/2790-0886_2026_74_3_35
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

The goal of this research is to create GridSense, a real-time digital platform that uses stream data processing and machine learning to forecast electricity consumption in Petropavlovsk city. The study tackles the issue of manual load planning in the North Kazakhstan regional power system, where dispatch services face considerable unpredictability due to significantly continental climate conditions, wind influence, heating demand, and noticeable daily consumption peaks. The primary goal was to develop an economically accessible, low-latency, and interpretable decision-support platform that could generate short-term forecasts up to 48 hours in advance. The work's methodological foundation consists of the creation of an event-driven software architecture based on Apache Kafka and system-structural modeling using Data Flow Diagram (DFD) and Integration Definition for Function Modeling (IDEF0). A Linear Regression model trained on climatic and temporal parameters, such as temperature, wind speed, humidity, present load, forecast horizon, and goal hour, is used by the forecasting module. A browser-based single page application, a Node.js and Express.js backend, a Kafka message broker, a Python machine learning consumer, a PostgreSQL database, and WebSocket real-time delivery are all integrated into the implemented platform. The model attained an operational forecasting accuracy of 88.8% with R² = 0.77 and Mean Absolute Error of 23.7 MW, according to experimental evaluation. With a mean response time of 8.9 ms, p95 latency of 12.1 ms, p99 latency of 18 ms, and no detected errors, load testing with Artillery showed robust system operation. The suggested approach serves as a foundation for intelligent energy planning systems in Kazakhstan and offers useful assistance for regional dispatch management.