Beyond two weeks: How AI is transforming Indian monsoon forecasting

Recent extreme rainfall and flooding in Assam, along with recurring flash floods and landslides across the Himalayan states, highlight India’s vulnerability to monsoon extremes. Forecasts cannot prevent these events, but reliable warnings three to four weeks ahead could give governments, water managers and communities valuable time to prepare.

Producing useful forecasts at such lead times remains one of meteorology’s hardest problems. Physics-based numerical models forecast weather remarkably well for about seven to 10 days, but their accuracy declines over longer horizons. As artificial intelligence models begin to match conventional models at shorter ranges, researchers are asking whether AI can also anticipate unusually wet, dry or hot conditions several weeks in advance.

For India, even a modest increase in useful lead time could improve decisions in agriculture, reservoir management and energy planning.

The forecasting gap

Meteorologists call the period from roughly two weeks to two months ahead the sub seasonal-to-seasonal, or S2S, range. Unlike a short-term weather forecast, an S2S outlook does not try to predict rain at a specific place and hour. It estimates whether an upcoming week is more likely than usual to be wetter, drier, hotter or colder across a region.

This range has traditionally been difficult to predict. By the third and fourth week, the atmosphere has lost much of the detailed memory of its initial state. Slower changing features, however, can retain useful information. These include ocean temperatures, soil moisture, snow cover and broader patterns in the land and atmosphere.

Our ongoing work examines whether AI can extract more information from these persistent signals. It builds on predictions from leading global weather models and learns to refine, interpret and optimally combine them.

Early experimental results from the 2026 monsoon are encouraging, particularly at three and four week lead times. Analysis of monsoon cases since 2022 also suggests that AI could extend useful forecast skill by roughly one to two weeks.

Why weeks three and four matter

Many monsoon-sensitive decisions must be made before reliable short-range forecasts become available. Farmers arrange seed and labour, reservoir managers decide whether to retain or release water, and power-system operators schedule maintenance and generation.

By the time a conventional forecast identifies a wet or dry spell with high confidence, some of these decisions may already be costly to reverse.

Weeks three and four occupy a valuable middle ground. They provide enough time to adjust plans while remaining close enough for the ocean, land and large-scale atmosphere to offer predictive signals. Extended-range forecasting centres therefore issue weekly regional outlooks rather than attempting to predict individual days a month ahead.

An S2S forecast is not a 28 day version of a smartphone weather forecast. A useful outlook might indicate an increased probability of a wetter-than-normal week over central India or sustained heat across a broad region. If reliable, even this level of information can support better preparation.

How AI can help

AI can identify subtle relationships across large volumes of weather and climate data. Some of these relationships may remain informative after individual weather systems can no longer be predicted precisely.

Ocean temperatures, soil moisture, snow conditions and large-scale tropical weather patterns evolve relatively slowly. They provide a form of memory that AI systems may be able to exploit.

Uncertainty nevertheless remains fundamental. At three- and four-week horizons, the goal is not a single certain prediction. It is to estimate the plausible outcomes and their probabilities.

AI can complement conventional meteorology by combining global model output with observations and historical data. Its value lies in extracting additional predictive information and translating it into clearer extended range guidance.

Agriculture

Agriculture is a natural application because farming decisions and monsoon variability operate over similar timescales.

Advance information about the onset and strength of rainfall can influence when and what farmers sow, how much land they cultivate and how much they invest in input. A forecast becomes most useful when it arrives before a decision becomes expensive to reverse.

Skilful forecasts for weeks three and four could eventually support advisories on sowing, irrigation, fertiliser application and preparation for prolonged wet or dry periods.

Energy and water

Monsoon conditions also shape India’s energy and water systems. Rainfall affects reservoir storage and hydropower, temperature influences electricity demand, and weather conditions determine wind and solar generation.

Reliable signals three or four weeks ahead could help operators assess a wider range of demand, generation and reservoir scenarios before committing to plans. Research on electricity demand, wind generation and hydropower already suggests that sub-seasonal information can add value at these horizons.

The aim is to give decision-makers earlier notice that the balance of probabilities is shifting towards an unusually wet, dry or hot period.

AI will not deliver precise weather forecasts a month in advance. It may, however, extract more information from the signals that persist. For India, extending useful monsoon forecast skill by even one or two weeks could make the difficult space between weather and seasonal forecasting substantially more valuable.

Sandeep Juneja is Director, and Manmeet Singh is Academic Visitor, Safexpress Centre for Data, Learning and Decision Sciences, Ashoka University, Delhi-NCR

Views express are personal and do not reflect the views of the university or THE WEEK.