A collaborative team from the High School of Economics and the Yandex School of Data Analysis is advancing a neural network project to forecast the El Niño climate pattern, leveraging the Yandex Cloud platform to support development and experimentation.
The enhanced algorithm aims to improve the accuracy of predicting shifts in the average surface temperature of the Pacific Ocean. These temperature changes are a primary driver of natural disasters in several regions. The current model offers forecasts up to 1.5 years in advance, with a longer horizon of two years envisioned as the research progresses and data streams widen. This extended lead time would give communities and authorities more space to prepare for potential impacts on weather, agriculture, and infrastructure.
The neural network focuses on the equatorial belt of the Pacific and tracks how average temperatures there fluctuate during El Niño. When El Niño intensifies, the equatorial zone tends to warm beyond typical levels. Conversely, La Niña represents a cooling phase, and the two phenomena typically alternate every two to seven years. These cycles have broad consequences for weather patterns around the globe, contributing to increased fire risk, droughts, floods, and crop shortages in vulnerable regions.
Researchers trained the models on thousands of temperature maps drawn from both synthetic datasets and real observations spanning from the year 1800 to today. In addition to standard predictive methods, the team tests cutting-edge architecture options, such as Autoformer, to handle time-series data with higher fidelity. This approach enables the model to extract meaningful signals from long sequences of temperature maps, improving forecast quality and reliability. The preprocessing stage relies on Yandex DataSphere ML development services, which provide a full toolkit and scalable cloud resources that support an end-to-end machine learning workflow—from data preparation to model training and validation.
“In projects like the El Niño study, rapid and flexible access to services matters for trying different machine learning configurations,” noted a representative from the initiative. “Each test with a new architecture is aimed at delivering earlier and more accurate predictions.”
Prior discussions highlighted a UN context, where a warming El Niño signal could potentially override La Niña conditions in upcoming months, underscoring the global urgency of improved monitoring and forecasting capabilities.