CastScience

Science Info and Calculations

Water temperature

How the app projects the lake temperature for the days ahead, and how wide the uncertainty is.

In plain words

Water holds a lot of heat, so a lake changes temperature slowly. Day by day it moves toward the air temperature: when the air is colder than the water the lake cools, and the larger the difference, the faster it cools. That is Newton's law of cooling. Wind speeds it up: it mixes the water and carries heat away from the surface faster, so the app lets the rate grow with wind speed.

The model is fitted to Lake St. Clair itself: 5,972 stretches of 7 days in 23 years between 1995 and 2026, GLSEA lake temperature against the weather at Selfridge, with its own coefficients for each calendar month. Before a plan is built, it carries the latest measured lake temperature forward day by day to the plan's date.

The heat-balance model

One step per day
Tn+1=Tn+(k0+kwUn)(Tair,n−Tn)T_{n+1} = T_n + \left(k_0 + k_w U_n\right)\left(T_{\text{air},n} - T_n\right)

TnT_n: lake-average water temperature on morning nn, F. TairT_{\text{air}}: the day's mean air temperature, (high + low) / 2, at Selfridge, F. UU: the day's mean wind at the shore station, mph. k0k_0 and kwk_w are those of the day's calendar month (lake record; fitted 2026-10-02):

Monthk0, per daykw, per day per mphk at 8 mph, per day95% interval of k7-day stretches
Jan0.00000.00390.03120.026-0.04294
Feb0.02660.00220.04420.020-0.05592
Mar0.05120.00070.05680.037-0.080349
Apr0.06840.00000.06840.055-0.077585
May0.10230.00000.10230.086-0.119668
Jun0.08940.00450.12540.112-0.142642
Jul0.00000.01020.08160.064-0.097701
Aug0.00000.00750.06000.046-0.074688
Sep0.00000.01050.08400.074-0.091629
Oct0.01070.00840.07790.072-0.085640
Nov0.00000.00680.05440.050-0.064602
Dec0.00000.00470.03760.030-0.046282

The fit minimizes the error of 7-day projections, the length of the forecast that drives the projection, with k0≥0k_0 \ge 0 and kw≥0k_w \ge 0: the exchange cannot run backwards. A zero k0k_0 means the data show no exchange beyond what the wind term gives. The 95% interval comes from refitting on years drawn at random with repeats (bootstrap by year). A day without a date uses the all-season pair k0=0.0472,kw=0.0032k_0 = 0.0472, k_w = 0.0032, from the same fit with one pair for every month.

In Oct with a wind of 8 mph, k=0.0107+0.0084×8=0.0779k = 0.0107 + 0.0084 \times 8 = 0.0779 per day: the lake closes 7.8% of the gap between water and air each day. In steady weather half the gap is gone after ln⁡0.5/ln⁡(1−k)=8.5\ln 0.5 / \ln(1 - k) = 8.5 days. In continuous form this is dT/dt=k (Tair−T)dT/dt = k\,(T_{\text{air}} - T), with the solution T(t)=Tair+(T0−Tair) e−ktT(t) = T_{\text{air}} + (T_0 - T_{\text{air}})\, e^{-kt} and a time constant 1/k=12.81/k = 12.8 days.

The same form, relaxation toward an equilibrium temperature with an exchange coefficient that grows with wind speed, is the equilibrium-temperature method of Edinger, Duttweiler and Geyer (1968). The app uses the daily mean air temperature in place of the equilibrium temperature Working rule: a simplification; the monthly fit absorbs part of the difference, which is why kk is largest in May and June, when the sun warms the lake faster than the air alone would.

Weather for each day

  • Days already past: observed highs, lows and wind at Selfridge.
  • The next 7 days: the National Weather Service forecast for the shore point.
  • After day 7: NOAA 1991-2020 daily normal high and low for Selfridge (station USW00014804) and a wind of 8 mph Starting value: near the October mean at Selfridge (Wind over water). The app labels these days as climatology-driven.

Sensitivity: an error of 1.5 F in the air temperature moves the projection by k×1.5=0.12k \times 1.5 = 0.12 F on the first day and 1.5 (1−(1−k)7)=0.651.5\,(1 - (1 - k)^7) = 0.65 F after 7 days: the error grows more slowly than 7 times the first day's because the model keeps closing the gap.

Source: reference/python/projection.py and wx_analysis.py. Engine: projectWaterTemp in packages/engine/src/waterTemp.ts.

Golden run

The trip analysis ran the model from 64.346 F on 2026-09-29 with the NWS forecast and then Selfridge normals. The reference script runs the same days with the monthly coefficients (reference/python/heat_balance_monthly.py), and the engine reproduces it; largest difference 0.0005 F against a tolerance of 0.01 F (computed now). The trip analysis's own coefficients gave 62.14 F after Oct 6 where the monthly ones give 61.53 F.

Morning afterReference (Python)Engine
2026-10-0562.422 F62.422 F
2026-10-0761.180 F61.180 F
2026-10-0960.152 F60.152 F
2026-10-1159.159 F59.159 F

The 80% band

The heat balance gives the central value; history gives the spread. Across 31 years (GLSEA 1995-2025) the Oct 7-11 lake temperature averaged 60.9 F with a standard deviation of 3.9 F. Knowing the temperature on Sept 29 narrows that, by a straight-line fit (least squares) over the same years:

Oct 7-11 mean lake temperature vs Sept 29 lake temperature, GLSEA 1995-2025 (31 years)
TOct 7-11=6.55+0.844 TSept 29T_{\text{Oct 7-11}} = 6.55 + 0.844\, T_{\text{Sept 29}}
r=0.71,s=2.79 Fr = 0.71, \qquad s = 2.79\ \text{F}

The 80% band is the central value ±1.28 s\pm 1.28\, s; 1.281.28 is the 90th percentile of the standard normal distribution, so 10% of outcomes fall below the band and 10% above it. Here 1.28×2.79=3.61.28 \times 2.79 = 3.6 F. For 2026, with 64.35 F on Sept 29: 60.9±3.660.9 \pm 3.6 F (reference run: 57.3-64.5 F).

The app does the same for every day: the 1995-2025 climatology (mean, 10th and 90th percentile for each day of the year) and the regression slope and residual SD for each month and each lag of 1-30 days, computed from GLSEA when the data are loaded. The Water conditions screen shows the projection inside this band.

Source: data/golden/engine_golden.json (regression); reference/python/temps.py. Engine: linearFit, band80, Z80 in waterTemp.ts.

Measured error

Every ice-free day of 1995-2026 with usable Selfridge weather was projected 7 days ahead with the weather that happened and compared with GLSEA 7 days later (5,972 cases). Root-mean-square error: 1.63 F with the monthly coefficients, 1.81 F with the October-only coefficients used before the 2026-10-02 calibration (k0 0.0135, kw 0.0058); 80% of the errors lie between -2.4 and +1.6 F. A real plan uses forecast weather, whose errors add to these. Fitted on even years and tested on odd years, and the reverse, the monthly coefficients gave 1.67 F against 1.81 F.

The October-only coefficients ran warm in strong wind: a hindcast of Sept 18-29, 2026 was up to 2.7 F warm during a three-day northeast blow. With the monthly coefficients the same hindcast is at most 1.8 F warm. Over all windy weeks (mean wind 12 mph or more) the 7-day error averages -0.06 F. More in Validation.

Source: reports/calibration-1.8-20261002.md; reports/calibration-latest.json (scripts/calibration/watertemp.ts).

Calculator

Water temperature projection

Runs the app's engine in this browser: projectWaterTemp (packages/engine/src/waterTemp.ts).

Lake St. Clair coefficients (lake record), by the month of each day: September k0 = 0 per day, kw = 0.0105 per day per mph; October k0 = 0.0107 per day, kw = 0.0084 per day per mph. Each row is one day of weather; the last column is the lake temperature the next morning. Dates follow the golden run.

DayAir high, FAir low, FWind, mphLake next morning, F
2026-09-3064.27
2026-10-0164.67
2026-10-0264.30
2026-10-0363.57
2026-10-0463.18
2026-10-0562.42
2026-10-0661.53
After 7 days61.53 F
Change-2.81 F
Gap closed on day 18.4%
Table of values
Lake temperature each morning and daily mean air temperature
Days from the startLake, each morningAir, daily mean
day 064.3 F63.4 F
day 164.3 F69.5 F
day 264.7 F61.0 F
day 364.3 F55.0 F
day 463.6 F58.5 F
day 563.2 F53.5 F
day 662.4 F51.0 F
day 761.5 F-

Constants

ConstantValueIn the codeBasis
Exchange rates k0 and kw by monthtable aboveheat_monthly (lake record)Fit, 7-day projections, GLSEA and Selfridge 1995-2026 (reference/python/heat_balance_monthly.py, 2026-10-02)
All-season k0 and kw0.0472 per day, 0.0032 per day per mphheat_k0_per_day, heat_kw_per_day_mph (lake record)Same fit, one pair for every month; for a day without a date
Wind past the forecast8 mphengine.default_day_wind_mph (lake record)Starting value: near the October mean
Normal-distribution z for 80%1.28Z80 (waterTemp.ts)90th percentile of the standard normal (1.2816), rounded
Regression for Oct 7-116.55 + 0.844 x; s = 2.79 Fregression (data/golden)GLSEA 1995-2025, 31 years
Normals stationUSW00014804normals_station (lake record)NOAA NCEI 1991-2020 daily normals

Sources

  • NOAA GLERL, Great Lakes Surface Environmental Analysis (GLSEA), daily lake-average temperature, 1995-2026.
  • Iowa Environmental Mesonet, Selfridge ANGB (KMTC) daily summaries; National Weather Service forecasts; NOAA NCEI U.S. Climate Normals 1991-2020.
  • Edinger, J.E., Duttweiler, D.W. and Geyer, J.C. (1968). The response of water temperatures to meteorological conditions. Water Resources Research 4: 1137-1143.
  • Links: References.