Where Can AI Help Flood Early Warning? Lessons from Bangkok's 50-District Flood, September 2026

1. What happened: who knew what, and when

This was not rain that was violent by the hour. It was rain that did not stop for more than 30 hours. The cloud mass rotated in place over the city, a pattern Thai forecasters call "rain stuck in the sky". Accumulated totals therefore ran three to four times what the drainage system can handle, and every evening a high tide kept the canals from emptying into the Chao Phraya River. The numbers worth remembering:

• Bangkok's drainage handles roughly 60 mm of rain, as the Governor has said before; some sources put it at 50–80. When rain runs across days the real limit is total pumping of 1,000–1,200 cubic metres per second, which over the city's 1,569 square kilometres clears about 66 mm of rain a day
• From 16:00 on the 24th to the morning of the 26th, Min Buri accumulated 274.5 mm and Khlong Sam Wa 273.0 mm; by afternoon the Governor said several eastern districts had passed 300 mm
• The Thesaban Songkhro gauge in Chatuchak District read 210 mm in the Drainage Department's morning report on the 26th. The Governor estimated about 223 million cubic metres of rainwater fell on the east bank in a day and a half
• Rain fell almost continuously for nearly 48 hours, with pumps on both banks running at full capacity throughout
• High tide at Bangkok reached about 3 metres between 18:30 and 19:30 every evening that week, with a full moon on 26 September

The table lines up, day by day, what agencies knew and what residents actually received.

Date	What agencies knew and said	What Bangkok residents actually received
21 Sep	The Office of the National Water Resources (ONWR) issued Announcement 5/2569 telling Bangkok, Nonthaburi, Pathum Thani and Samut Prakan to watch for urban flooding on 22–26 September. An independent forecasting page posted that rain would fall day and night on 24–27 September, 120 hours ahead, and told people to move belongings upstairs	General news
22 Sep	The Thai Meteorological Department (TMD) issued Bulletin No. 1: heavy to very heavy rain on 23–27 September, with Bangkok in the "very heavy" band on 24–26. The same day it stated that the 120-hour post did not match its forecast data	A general warning, and a message not to panic
23 Sep	The Department of Disaster Prevention and Mitigation (DDPM) ordered 70 provinces and Bangkok to ready teams, machinery and shelters. The BMA drew down water in 1,980 canals. The Governor said localised downpours have to be watched day by day	"Allow extra travel time" and download the alert app
24 Sep	Thailand's first Cell Broadcast of the event went out at 10:01 to Trat Province. Continuous rain over Bangkok began at 16:00	No message to Bangkok phones yet
25 Sep	Min Buri recorded 101.5 mm in 24 hours and 15 roads were under water. Three districts were declared disaster areas. DDPM reported 12 provinces already flooded, including Sa Kaeo, Chachoengsao and Prachin Buri	Street-by-street flood reports in the news
26 Sep	Accumulated rain at Min Buri reached 274.5 mm by morning and passed 300 mm by afternoon. Cell Broadcast reached Bangkok at 09:22: canal levels critical, move belongings to higher ground. All 50 districts were declared disaster areas. The government held a command meeting in the afternoon. One person died from electrocution in floodwater	The first warning that said what to do, about 41 hours after continuous rain began

Notice that every piece was in place by the 22nd: a forecast, a 300 mm flood simulation the BMA had already built, a tide table known a year in advance, a system that can push a message to every phone, and a canal drawdown plan. What did not exist was the single sentence that joins those pieces together: "flooding is likely, finish these things before the evening of the 24th". The rest of this article looks at where AI can help join them.

2. The gap is not the forecast. It is four seams

Read the timeline and the forecast did not miss. TMD put Bangkok in the "very heavy" band, which by its own definition means more than 90 mm in 24 hours, for three consecutive days. That already exceeds the roughly 66 mm a day the city's pumps can clear, every single day. The problem sits at four seams between the forecast and action.

Seam 1: data scattered across agencies and formats. TMD issues text bulletins. ONWR issues documents. The BMA has canal-level telemetry and two radar stations. The Navy publishes the tide table. The Royal Irrigation Department has dam discharge rates. DDPM holds the Cell Broadcast button. Each piece is correct. No single screen shows them together.

Seam 2: the forecast is never translated into impact. "Very heavy rain" is a meteorologist's term. A resident does not know how deep the street outside will flood at 90 mm, or for how long. Yet the BMA already has a 300 mm flood simulation. Feeding each new forecast into that model is a job nobody is assigned to do.

Seam 3: the warning does not say who must do what, by when. On the 22nd and 23rd residents were told to "stay vigilant" and "allow extra travel time". On the 26th they were told to "move belongings to higher ground". The second message is the one that should have arrived on the 23rd, with a deadline of the evening of the 24th.

Seam 4: handling information from outside the system. The 120-hour post used the phrase "mega-flood", which overstates the case by definition: northern inflow at the Chao Phraya Dam this year was 1,950 cubic metres per second against 3,721 in 2011. But the substance of the post, continuous day-and-night rain and moving belongings upstairs, was close to what happened. Under time pressure an agency has to choose between rejecting the whole post and separating the exaggerated word from the usable substance. The second option needs someone to read and compare within hours, which is exactly the kind of work AI does well.

3. Six places AI helps

The table follows the four seams. Each row says what AI does, what people still have to do, and whether a real-world example already exists. Points 1, 4, 5 and 6 can be done immediately without waiting for a large project.

Point	What AI does	What people still do	Real examples
1. Data fusion	Reads bulletins, documents, posts, tide tables and telemetry from every agency and produces one situation picture, refreshed every six hours	Decide which sources are trusted and agree data sharing between agencies	A large language model (LLM) can read mixed-format documents today. No new system needs to be built
2. Rain forecasting	Data-driven weather models now match or beat physics-based models on many measures and use roughly 1,000 times less energy per run, so they can run more often and produce more scenarios	Interpret the output, check it against local radar, and own the forecast	The European Centre for Medium-Range Weather Forecasts (ECMWF) has run its Artificial Intelligence Forecasting System (AIFS) operationally since 25 February 2025. Google DeepMind's GenCast beat ECMWF's ensemble system (ENS) on 97.2 percent of evaluation targets
3. Impact translation	Feeds forecast rain, canal levels and tide times into the flood model that already exists, and outputs a map of which districts flood, how deep, for how long, refreshed every forecast cycle	Validate the model against real events and own the canal-level data	Google Flood Hub does this for rivers in more than 150 countries, seven days ahead. It covers river flooding, not urban drainage flooding
4. Devil's advocate	Instructed to argue against the forecast team's conclusion using the same data, especially when the conclusion is milder than the model, and to separate exaggeration from substance in independent posts	Decide which side to believe, and sign	The same technique we used in our Five Forces article where the AI argued against its own scores. One score in five flipped
5. Plain-language warnings	Drafts warnings per audience, canal-side residents, drivers, employers, hospitals, the elderly, and per language, Thai, English, Myanmar, Khmer, all from one set of numbers, within minutes	Read, edit, sign, and take responsibility	Hong Kong uses Amber, Red and Black signals at 30, 50 and 70 mm per hour. Japan uses five levels, with "finish evacuating at Level 4"
6. Public Q&A	A chatbot answers "which roads are passable, where can I park, what number do I call" from official real-time data; CCTV image analysis reads water depth; complaints are triaged for district offices	Keep the source data correct and current	Wisesight social data: parking questions rose 127-fold on the 25th, emergency-number sharing rose 18-fold on the 26th

Three points deserve a closer look, because they are the ones people either assume AI cannot do or assume it does better than it really does.

Point 3, impact translation. No new model is needed. The BMA already has a 300 mm flood simulation, 24-hour canal telemetry and a tide table. What is missing is a person or a program that runs the three together every time TMD issues a new bulletin. That is work a program does better than a person, because it has to repeat every six hours and must never be forgotten.

Point 4, the devil's advocate. Every forecast team is pulled toward the milder reading, because a strong warning that does not materialise draws criticism. Having an AI argue "if the model says three straight days in the very-heavy band and the city can clear about 66 a day, why is the conclusion only stay vigilant?" does not mean the AI is right. It forces the team to write its reasoning down before deciding. The 120-hour post is the same case. An AI can compare the post against the official bulletin in minutes and list where they agree and where they differ. The public statement then changes from "this is not true" to "this is not a 2011-style mega-flood, but continuous rain is real, and here is what to prepare".

Point 5, plain-language warnings. Impact-based warnings are the approach the World Meteorological Organization (WMO) has promoted since 2015 and the core of its Early Warnings for All initiative, which aims to protect everyone on Earth with an early warning system by 2027. The next table compares how existing systems differ.

System	Criteria	What the public must do
Thailand, TMD bulletins	"Heavy" above 35 mm and "very heavy" above 90 mm per 24 hours, listed by region and province	"Beware of the danger of heavy and accumulated rain"
Thailand, Cell Broadcast since 2025	Sent when the local situation reaches a critical level, triggered manually by DDPM	The 26 September message: move belongings upstairs, avoid flooded routes
Hong Kong, three-colour rainstorm signals	Amber 30, Red 50, Black 70 mm per hour and likely to continue; can be issued below threshold when rain is prolonged	Black signal = schools, offices and courts stop citywide with no further order needed
Japan, five-level alerts, revised 28 May 2026	Level 4 = everyone in at-risk areas must evacuate. Level 5 = disaster under way, evacuation no longer safe	"Finish evacuating at Level 4" is a sentence the public can recite
WMO Early Warnings for All	Target: everyone on Earth protected by early warning by 2027. Today 119 countries, 60 percent, report a multi-hazard early warning system	A warning must state impact and required action, not just rainfall

4. The warning AI could have drafted on the evening of the 22nd

Below is a draft an AI could produce in minutes from the data available on the 22nd. A forecaster reads it, corrects the numbers and signs. The AI does not send it.

This draft differs from the warnings actually issued in four ways. It has numbers compared against the city's limit. It has a deadline. It lists actions. And it publishes the escalation trigger in advance. The last one matters most, because it makes escalation automatic from telemetry, with no need to wait for a 9 a.m. meeting on the day the city is already under water.

5. What AI cannot do, and what agencies must do themselves

Four things will not move with even the best technology unless a person decides.

• Accepting false-alarm risk. Closing schools or ordering work-from-home in advance is a policy decision. AI can supply the probability. A person has to choose whether to act at 60 percent or wait for 90
• Data-sharing agreements between agencies. If canal telemetry, radar, dam discharge and tide data do not flow together automatically, the AI has nothing to read. This is a matter of regulation and law, not technology
• Physical capacity. The city handles roughly 60 mm of rain. AI does not add pumps. Experts put the cost of upgrading to 100–150 mm above the cost of ten mass-transit lines
• Accountability. Every warning needs a signature. AI can draft it. The name on the notice has to be a person's

For an agency that wants to start, this table is split by time and money. The first row can be done with tools already in hand, with no new procurement.

Horizon	What to do	Prerequisite
Now 0–3 months	Have an LLM read every agency's bulletins and produce one situation summary every six hours. Draft warnings per audience and per language from one set of numbers. Add a devil's-advocate step before each bulletin. Run a chatbot on the BMA's live flood-point page	AI accounts for the team, and rules on who checks and who signs
Medium 6–12 months	Connect the 300 mm flood model to the forecast so it runs automatically every cycle. Set escalation triggers from telemetry and canal levels. Use CCTV image analysis to read water depth	Data-sharing agreement among the BMA, TMD, Royal Irrigation Department, Navy and DDPM
Major investment 1–3 years	AI radar nowcasting of the kind the Hong Kong Observatory runs. A national colour-coded warning system with numeric criteria and required actions tied to each colour	Budget, enabling legislation, and public drills

Conclusion

The Bangkok flood of 24–26 September 2026 was not a forecasting failure. The "very heavy" forecast was issued on the 22nd. What was missing were four seams: scattered data, a forecast never translated into impact, a warning that did not say who must do what by when, and the handling of information from outside the system. AI can close those seams at six points, and four of the six can be done today with tools already in hand. What AI cannot replace is the decision to accept false-alarm risk, the data-sharing agreement, and the name of the person who signs the warning.

A forecast has not helped anyone until it becomes a sentence that says who must do what, and by what time.

- Saeree ERP Team

References

• Thai Meteorological Department — Weather warnings: heavy to very heavy rain, 23–27 September 2026 (Bulletin No. 1 issued 22 September)
• The Nation — Bangkok among areas facing very heavy rain on September 24–26 (22 September 2026)
• Bangkokbiznews — TMD responds to the "mega-flood" rumour, confirms heavy monsoon rain 23–27 September (22 September 2026, Thai)
• Thai PBS — TMD defines "flood" versus "mega-flood" (23 September 2026, Thai)
• The Thaiger — The "120-hour mega-flood" post revisited, and DDPM's report of 12 flooded provinces (25 September 2026, Thai)
• National water situation summary, 22 September 2026, including ONWR Announcement 5/2569 (Thai)
• Thairath — DDPM tells 70 provinces and Bangkok to watch for flash floods 23–27 September (23 September 2026, Thai)
• Thairath — Governor Chadchart: localised downpours cannot be assessed in advance; 1,980 canals drawn down (23 September 2026, Thai)
• Spacebar — 30 Cell Broadcast messages in 53 hours across 15 provinces (26 September 2026, Thai)
• Thairath — DDPM sends Cell Broadcast to Bangkok: canal levels critical, 26 September 2026 at 09:22 (Thai)
• Bangkokbiznews — Bangkok rain passes 210 mm at the Chatuchak gauge (26 September 2026, Thai)
• THE STANDARD — Rain stuck over Bangkok: 50–80 mm drainage design capacity and the tide effect (Thai)
• Thai Post — Governor concedes 300 mm is hard to handle; the 300 mm simulation and 223 million cubic metres on the east bank (Thai)
• Thai PBS — Bangkok faces nearly 300 mm in 48 hours; continuous rain since 24 September 16:00; Min Buri 274.5, Khlong Sam Wa 273.0 (26 September 2026, Thai)
• Matichon — Governor Chadchart: Bangkok's drainage handles about 60 mm (Thai)
• Bangkok Post — All 50 Bangkok districts declared disaster areas (26 September 2026)
• Thai PBS — Government orders all agencies to respond; use Cell Broadcast in plain language (26 September 2026, Thai)
• Thairath — 2026 flood versus 2011: Chao Phraya Dam discharge 1,950 versus 3,721 cubic metres per second (Thai)
• Royal Thai Navy tide table for Bangkok, September 2026
• Wisesight — What social media said about the September 2026 flood: 126,524 messages (Thai)
• Daily News — Electrocution death in floodwater at Khlong Chan housing estate (26 September 2026, Thai)
• Public Relations Department — Cell Broadcast Service emergency alerts launched in 2025 (Thai)
• ECMWF — AIFS becomes operational, 25 February 2025
• Nature — Probabilistic weather forecasting with machine learning (GenCast), December 2024
• Google Research — Flood Forecasting and Flood Hub coverage
• Hong Kong Observatory — Rainstorm Warning System (Amber 30 / Red 50 / Black 70 mm per hour)
• Hong Kong Observatory — Developments and Prospects of AI in Rainfall Nowcasting
• Nippon.com — Five-Level System Simplifies Disaster Information Sharing in Japan
• WMO — Early Warnings for All shows real progress (2025)