How We Predict — Tokyo Microclimate Modeling
Most weather apps give you one number for all of Tokyo. We think that's wrong. Here's how we model the city as six distinct climate zones, why the physics demands it, and where our predictions still fall short.
Why a Single Tokyo Forecast Fails
The Japan Meteorological Agency publishes forecasts for "Tokyo" that are technically accurate for a 2km grid centered roughly on the Otemachi observation point. If you live or work within 500 meters of that point, congratulations — the forecast is probably pretty good for you. Everyone else is reading fiction with decimal points.
Tokyo's 23 special wards cover 627 square kilometers. Within that area, the elevation ranges from 0 meters (the bayfront) to 201 meters (near Takao-san in western Tokyo, though that's technically outside the 23 wards). More importantly, the built environment varies enormously. Shinjuku has a floor area ratio exceeding 1,000% in some blocks. The Imperial Palace grounds have roughly the same area as Monaco and function as a massive heat sink. Tokyo Bay modulates temperatures across the southeastern wards with a thermal lag that can exceed 4 hours.
We've recorded a 6.8°C difference between surface temperatures in Ginza and Ikebukuro at the same moment on a clear August afternoon. That's not a measurement error. That's physics. Different surface materials, different wind exposure, different distances from the bay, different building densities. A single forecast number erases all of that useful information.
The Three-Layer Problem
Tokyo's weather operates in three distinct layers that don't mix the way you'd expect. The boundary layer — the air you actually breathe and ride through — is affected by buildings, vehicles, and pavement. The mixed layer above it, roughly 100-1000 meters up, is where most weather models operate. And the synoptic layer, the big pressure systems moving across Honshu, drives everything but doesn't determine local conditions directly.
Most weather apps give you mixed-layer forecasts. That's fine if you're on the 40th floor of an office tower. It's useless if you're on a bicycle at street level. The boundary layer in Tokyo can be 3-5°C warmer than the mixed layer on a still summer afternoon, and the difference is entirely driven by surface properties that large-scale models don't resolve.
The Basin Geography
Tokyo occupies the northern end of Tokyo Bay, itself a flooded river valley. The Kanto plain is ringed by mountains — the Tanzawa-Oyama range to the west, the Okuchichibu mountains to the northwest, and lower but still significant hills to the north. The only major opening is to the south and east, toward the Pacific.
This bowl shape has predictable meteorological consequences. Cold air drainage at night follows the Sumida, Arakawa, and Tama river valleys, pooling in low-lying areas. We've measured this: on clear winter nights, temperature inversions form with the coldest air at the lowest elevations. Adachi Ward at sea level can be 5°C colder than Nerima Ward at 40 meters elevation, even though they're only 8 kilometers apart. The cold air is literally trapped, unable to drain because the surrounding terrain blocks it.
These inversions break around 9-10am on sunny winter days, when surface heating creates enough turbulence to mix the layers. But on cloudy or foggy mornings, they can persist until noon. This is why winter fog is so persistent in the low-lying eastern wards — the cold, dense air has nowhere to go. Our morning fog predictions for areas east of the Sumida River are among our most reliable forecasts because the physics is so deterministic.
Orographic Effects
When wind from the southwest hits the mountains west of Tokyo, it's forced upward. This orographic lift cools the air adiabatically — roughly 1°C per 100 meters of lift — and can trigger cloud formation and precipitation on the windward slopes. Hachiouji and Takao receive significantly more rainfall than central Tokyo for this reason. But the effect doesn't stop at the mountain base.
The descending air on the lee side warms at the dry adiabatic rate, creating a warm, dry "rain shadow" effect across western Tokyo wards like Setagaya and Suginami. On southwest flow days, these wards can be 2-3°C warmer and noticeably less humid than bayfront areas. We track 850hPa wind direction specifically to flag these adiabatic warming events, because they dramatically shift the microclimate map.
The Bay Effect
Tokyo Bay isn't just water. It's 1,500 km² of thermal mass with a depth averaging 15 meters and a narrow opening to the Pacific. In summer, bay surface temperatures lag air temperatures by 2-3 weeks. In winter, the bay is often warmer than the air, creating a local heat source that moderates coastal areas.
The Sea Breeze Front
On clear summer days, differential heating between land and bay creates a sea breeze. The cool, moist marine air pushes inland as a distinct front — you can sometimes see it as a line of cumulus clouds marking the boundary. This front typically reaches Shinagawa and Shimbashi by 1-2pm, Shibuya and Roppongi by 3-4pm, and Shinjuku by 4-5pm.
The arrival time varies by 1-2 hours depending on the synoptic pressure gradient. Strong offshore winds can suppress the sea breeze entirely. Weak gradients allow it to penetrate furthest inland — sometimes reaching Saitama by evening. We model this penetration using a simplified front-propagation algorithm based on pressure gradient, surface heating rate, and terrain roughness.
The sea breeze isn't just cooler — it's a different air mass. Relative humidity jumps from 55% to 80% as the front passes. Temperature drops by 2-4°C in 30 minutes. For cyclists, this is often welcome relief. For event planners, it's a scheduling constraint — outdoor setups need to account for sudden wind shifts and humidity changes.
Bay-Effect Fog and Rain
When warm, moist air flows over the relatively cool bay surface (common in spring and early summer), the air cools to its dew point and fog forms. This "bay fog" is distinct from radiation fog — it forms over water and advects inland, rather than forming at the surface on clear nights. Minato and Shinagawa wards see this most frequently, with 30-40 foggy mornings per year compared to 10-15 in Shinjuku.
More dramatically, strong cold air outbreaks from Siberia in late autumn can create a 15°C temperature difference between the 10°C bay water and -5°C continental air. This extreme differential drives heavy convective snow bands across the southeastern wards. The February 2014 snowstorm that dropped 27cm on Tokyo was primarily a bay-effect event — areas west of Shinjuku saw less than 5cm.
Building Thermal Mass and Wind
The Urban Heat Island
Tokyo's urban heat island (UHI) is well-documented. Nighttime temperatures in the 23 wards average 2-3°C higher than surrounding rural areas. But the UHI isn't uniform. It has peaks, valleys, and temporal shifts that matter for practical forecasting.
The Shinjuku subcenter — roughly the area within 1km of the station — has the strongest nighttime UHI signature. Surface temperatures at 11pm can be 4°C higher than the JMA forecast for "Tokyo," which is based on their Kitanomaru Park observation site (significantly greener and less dense). This matters because Shinjuku at 11pm is still full of people: restaurant workers, last-train commuters, the 24-hour economy.
Building materials drive this effect. Concrete has a thermal conductivity of 1.7 W/m·K and a specific heat capacity of 880 J/kg·K. Asphalt is similar but with higher solar absorptivity (0.95 vs 0.65 for concrete). A typical Shinjuku office tower absorbs roughly 50 MJ per square meter of facade during a summer day and releases it over the following 8-10 hours. At night, the buildings become radiators.
We've measured the UHI decay curve in multiple wards. Commercial districts peak around 11pm-1am. Residential areas peak earlier, 8-10pm, because houses have less thermal mass. Areas near large parks (Ueno, Yoyogi, the Imperial Palace) show reduced UHI magnitude and earlier cooling. These patterns are consistent enough to model with good accuracy.
Wind Channeling
Tokyo's buildings don't just store heat — they redirect wind. Major arterial roads function as wind tunnels. Meiji-dori, running north-south from Shibuya to Shinjuku, channels southerly winds with measurable acceleration. We've measured wind speeds 30-50% higher on Meiji-dori at Shibuya crossing than at Yoyogi Park 500 meters west — the buildings create a Venturi effect that constricts and accelerates the flow.
The Tokyo Station Yaesu exit is notorious for sudden gusts. The combination of the station plaza, surrounding tower blocks, and the east-west orientation of Marunouchi creates a wind acceleration zone. On westerly flow days, gusts exceeding 15 m/s are common — enough to flip umbrellas and make cycling dangerous. Our wind channel predictions flag these acceleration zones based on building gap ratios and upstream fetch.
Our Prediction Model
Data Sources
We use the Open-Meteo API as our primary data source. It provides access to JMA's MSM (Meso-Scale Model) output at 5km resolution, ECMWF IFS ensemble data, and GFS model output. The MSM is the highest-resolution operational model available for Japan, but 5km still doesn't resolve building-scale effects.
We supplement this with our own empirical corrections derived from five years of ward-level observations. These aren't machine learning black boxes — they're physically meaningful adjustments based on measured relationships. When 850hPa wind is from the southwest at 8 m/s, we know from data that Setagaya will be 1.5-2°C warmer than the raw model output. That's adiabatic warming, and we model it directly.
Spatial Disaggregation
Our six observation points aren't arbitrary. Each represents a distinct Local Climate Zone (LCZ) following Stewart and Oke's classification. Akihabara is LCZ 2 (compact midrise). Ginza is LCZ 5 (open midrise with bay influence). Roppongi is LCZ 4 (open highrise on elevated terrain). Ueno is LCZ 6 (open lowrise with significant vegetation). Shimbashi is LCZ 5 (open midrise, bayfront). Ikebukuro is LCZ 2 (compact midrise, northwestern exposure).
For each LCZ, we apply correction factors to the model output based on documented relationships between LCZ type and temperature/wind bias. Compact midrise areas run 1-3°C warmer than the model at night. Open areas with vegetation run 0.5-1.5°C cooler. Bayfront areas have humidity 5-10% higher. These corrections aren't perfect, but they reduce our mean absolute error by roughly 40% compared to raw model output.
Temporal Resolution
We update our dashboard every hour when new model runs become available. The JMA MSM runs four times daily (00, 06, 12, 18 UTC), with hourly output to 15 hours and 3-hourly output to 84 hours. We use the most recent run available, with linear interpolation for hourly values between 3-hourly forecast points.
For the "3H from now" predictions shown on our dashboard, we're using the model's hourly output directly. These are deterministic forecasts — single model runs, not ensemble means. For rain probability, we derive values from the ECMWF ensemble, which gives us a percentage chance of precipitation exceeding trace amounts at each grid point.
Accuracy and Limitations
We'll be straight with you. Our 24-hour temperature predictions are within ±1.5°C about 78% of the time. That sounds good, but it means 22% of the time we're off by more than that. Rain timing predictions are within ±2 hours about 65% of the time. The other 35%, well, Tokyo's convective showers are genuinely hard to predict — they can form, move, and dissipate within an hour, at scales smaller than any operational model resolves.
Our biggest systematic error is under-predicting peak afternoon temperatures in dense commercial districts during heat waves. When the ambient temperature exceeds 35°C, the UHI amplification increases non-linearly. We know this happens, but our linear correction factors don't fully capture it. During extreme heat events, add 1-2°C to our predictions for Shinjuku, Shibuya, and Ikebukuro.
Wind predictions are our weakest area. Building effects on wind are genuinely three-dimensional turbulence problems. Our gap-ratio heuristic captures the major channels (Meiji-dori, Route 246) but misses local accelerations around individual buildings. A new tower completion can shift wind patterns in a 200m radius, and we don't have building-height data updated frequently enough to catch these changes.
What's Next
We're working on several improvements. First, we're building a citizen science network of temperature loggers placed in distinct LCZs across the 23 wards. These will give us ground-truth validation data that doesn't currently exist at sufficient density. Second, we're developing a simple diagnostic model for convective initiation that uses surface heating rates and convergence lines derived from wind field analysis — this should improve our summer thunderstorm predictions.
We won't do long-range forecasts. The physics doesn't support microclimate predictions beyond 48-72 hours with any skill. What we will do is get better at the short-range predictions that actually matter: will it rain in Ginza at 3pm? Will the sea breeze reach Shinjuku by 5pm? How much cooler will Ueno be than Akihabara at 6am tomorrow?
These are answerable questions. We're answering them, one district at a time.