Coding and AI classes in Wolfsburg

import csv

temps = {}
with open("wolfsburg_daily.csv", encoding="utf-8") as f:
    for row in csv.DictReader(f):
        if row["mean_temp"] != "-999":
            year = int(row["date"][:4])
            temps.setdefault(year, []).append(float(row["mean_temp"]))

yearly = {}
for year, values in temps.items():
    if len(values) >= 365:
        yearly[year] = sum(values) / len(values)

print("Complete years:", len(yearly), " left out:", sorted(set(temps) - set(yearly)))

Three years drop out. 1981 starts in April and 2024 has a single day, which makes sense. But why 2019? The station's own metadata answers it: the daily means are missing from 3 to 12 June 2019 and again from 14 to 26 June 2019. A year without most of June would look colder than it was, so leaving it out is the honest choice.

train = [(y, t) for y, t in yearly.items() if y <= 2002]
test = [(y, t) for y, t in yearly.items() if y > 2002]

n = len(train)
mean_year = sum(y for y, t in train) / n
mean_temp = sum(t for y, t in train) / n
slope = sum((y - mean_year) * (t - mean_temp) for y, t in train) / sum((y - mean_year) ** 2 for y, t in train)

def line(year):
    return mean_temp + slope * (year - mean_year)

flat_error = sum(abs(t - mean_temp) for y, t in test) / len(test)
line_error = sum(abs(t - line(y)) for y, t in test) / len(test)
print("Learned slope:", round(slope * 10, 2), "degrees per decade")
print("Flat average, error on later years:", round(flat_error, 2))
print("Straight line, error on later years:", round(line_error, 2))

The model learned that, from 1982 to 2002, Wolfsburg's yearly mean rose by about 0.38 degrees per decade. On the twenty later years it never saw, the line is off by 0.48 degrees on average, while the flat old average is off by 0.85. A learned model beating a fixed guess on unseen data is the basic test every AI system has to pass.

for year in [2010, 2018, 2023]:
    print(year, "line says", round(line(year), 2), "station measured", round(yearly[year], 2))

Now the twist. The station's geography file shows that it has not always stood in the same place. Since 1 November 1990 it was at 70 metres; from 1 September 2010 it stands at 81.5 metres, at a new position. The model knows nothing about that. It sees one column of numbers called "Wolfsburg" and assumes they all mean the same thing. Whether the move changed the readings, and by how much, cannot be learned from this file alone. A careful data scientist writes that down next to every chart.

Data: Deutscher Wetterdienst, Climate Data Center, daily climate data and station metadata, station 05676 Wolfsburg (Südwest), licence CC BY 4.0, downloaded 3 November 2026. The yearly means, the line, its slope and its errors are our own calculations. We print no forecast beyond 2023: a line learned from the past is not a promise about the future.

1. Ages 6 to 9, no screen. Measure a plant's height every day for a week, draw the dots, and lay a ruler through them. Where does the ruler say it will be on day 10? Then measure on day 10.
2. Ages 10 to 12, Python. Run Step 1 and print the warmest and coldest complete year. Find 2010 and guess why it might stand out, then check the station table.
3. Teenagers. Fit the line on 1991 to 2009 only, the years at the middle site, and compare its slope with ours. What does the difference tell you, and what can it not tell you?
4. Adults. Replace the yearly mean with the yearly count of days above 25 degrees, using the max_temp column. Fit and test the same way, and write one sentence a manager could act on.