How to Use Feed-Forward Neural Networks to Predict Mumbai Cricket Weather

Weather forecasting for a cricket venue is a focused machine-learning problem, not simply a matter of training a neural network on a spreadsheet. At Mumbai’s Wankhede Stadium, short-duration rain, high humidity, sea-breeze effects, heat, and monsoon seasonality can change playing conditions within hours. A useful system must therefore predict a defined outcome at a defined time horizon and communicate uncertainty to match officials.

This guide explains how to use feed forward neural networks to predict weather in Mumbai Cricket Stadium, with an emphasis on reproducible data, leakage-free evaluation, and decisions such as rain alerts, ground-staff readiness, or likely playing-condition windows.

Define the prediction task first

Avoid starting with “predict the weather.” Choose one target and forecast horizon:

Rain occurrence: Will measurable rain occur at the venue in the next one, three, or six hours?
Rain intensity: What rainfall amount is expected over the next hour?
Temperature: What will the air temperature be at a specific match interval?
Humidity or wind: What conditions will affect player comfort, ball movement, or covers?
Operational outcome: Is there a high probability of a rain interruption during the scheduled innings?

For a first project, binary rain classification is often more useful than predicting every atmospheric variable. Define the label precisely—for example, rainfall above 0.1 mm in the following hour—and align every input with information that would genuinely have been available at prediction time.

A venue model should also distinguish weather at the stadium from a city-wide forecast. Mumbai’s coastal geography means observations from a distant station may not represent conditions at Wankhede. Use the closest reliable station, radar or satellite products where licensing permits, and venue-level observations when available.

Assemble Mumbai-specific data

Build a timestamped table at a consistent interval, such as 10 or 15 minutes. Potential inputs include:

Temperature, relative humidity, pressure, wind speed, wind direction, visibility, and rainfall.
Cloud cover, dew point, solar radiation, and wet-bulb temperature.
Recent rainfall totals over 15 minutes, one hour, three hours, and 24 hours.
Hour, day of year, month, and whether the period falls within Mumbai’s southwest monsoon.
Forecast variables from an approved weather API or the India Meteorological Department where access and usage terms allow.
Radar-derived precipitation distance or satellite cloud indicators, if available.
Match start time, innings window, scheduled breaks, and venue coordinates.

Store the source, retrieval time, unit, and quality flag for every observation. Historical weather APIs can revise data, so preserve raw responses rather than only a cleaned table. If you are building a broader forecasting workflow, the principles in Implementing Scalable ML Pipelines for Predictive Analytics are useful for versioning, monitoring, and reproducible retraining.

Engineer features without leaking the future

A feed-forward network does not understand time automatically. Convert temporal context into explicit features:

Lagged observations: rainfall, humidity, pressure, and wind from earlier intervals.
Rolling statistics: recent rainfall sums, humidity averages, pressure change, and wind variability.
Cyclical time encoding: represent hour and day-of-year with sine and cosine columns rather than treating 23:00 and 00:00 as far apart.
Monsoon indicators and interaction terms, such as humidity combined with recent pressure decline.
Forecast lead time and the age of each weather observation.

Do not use an observation recorded after the prediction timestamp, a corrected historical value unavailable at inference time, or a “final match status” field derived from the target. These are common sources of leakage and can make an apparently accurate model fail in deployment.

Categorical fields such as wind direction can be encoded as sine and cosine components. Scale continuous variables using statistics calculated on the training period only. Retain the fitted scaler so production data receives exactly the same transformation.

Build a practical feed-forward network

A compact multilayer perceptron is usually a sensible baseline. It can learn nonlinear relationships among recent rainfall, humidity, pressure trends, and seasonal variables without the operational complexity of a sequence model. Compare it with a persistence forecast, logistic regression, random forest, or gradient-boosted trees; a neural network is valuable only if it improves a meaningful metric.

import tensorflow as tf
from tensorflow import keras

model = keras.Sequential([
    keras.layers.Input(shape=(X_train.shape[1],)),
    keras.layers.Dense(64, activation="relu"),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(32, activation="relu"),
    keras.layers.Dense(1, activation="sigmoid")
])

model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=1e-3),
    loss="binary_crossentropy",
    metrics=[keras.metrics.AUC(name="auc"), "accuracy"]
)

callbacks = [keras.callbacks.EarlyStopping(
    monitor="val_auc", mode="max", patience=10, restore_best_weights=True
)]

Use sigmoid output and binary cross-entropy for rain/no-rain classification. For rainfall amount, use a linear output and a loss such as mean absolute error; rainfall is highly skewed, so consider modelling log1p(rainfall) or using a two-stage occurrence-plus-amount design. Customizable Neural Network Architectures for Beginners provides a useful introduction to changing layer sizes, regularisation, and output design.

Train and evaluate by time, not random rows

Randomly splitting adjacent weather observations can place near-duplicates in both training and test sets. Instead, use chronological blocks—for example, older seasons for training, a later period for validation, and the most recent monsoon season for testing. A rolling-origin evaluation is stronger: train on an expanding historical window and test on the next period.

For rain classification, report:

Precision, recall, F1 score, and area under the precision-recall curve.
Brier score and calibration plots for probability quality.
Confusion matrices at operational thresholds such as 30%, 50%, and 70%.
Performance separately during monsoon, non-monsoon, daytime, night matches, and heavy-rain events.

Accuracy alone is misleading when no-rain intervals dominate. A model that rarely predicts rain may achieve high accuracy while missing the interruptions that matter most. Select the threshold based on the cost of false alarms versus missed rain. Ground staff may prefer high recall, while a public alert system may require stronger precision.

Produce match-ready predictions

At inference time, create the same feature vector used during training, apply the saved scaler, and generate a probability rather than only a label:

probability = float(model.predict(X_new, verbose=0)[0][0])

if probability >= 0.70:
    status = "High rain risk"
elif probability >= 0.40:
    status = "Moderate rain risk"
else:
    status = "Lower rain risk"

Display the forecast with its horizon, last data update, confidence or calibration information, and the latest observed conditions. Refresh frequently before and during the match, but do not silently overwrite historical predictions. Logging predictions and outcomes lets you identify seasonal drift and compare the model with official forecasts.

A neural network should support—not replace—official meteorological guidance. Treat severe-weather warnings and operational safety decisions as higher-priority inputs. For production systems, add missing-data checks, API failure handling, outlier alerts, and a fallback forecast.

Common failure modes and next steps

The largest risks are sparse venue observations, inconsistent station locations, imbalanced labels, monsoon regime changes, and overfitting to one tournament or season. Begin with a transparent baseline, document every feature, and retrain only after monitoring confirms that new data improves out-of-sample performance.

If the feed-forward model cannot capture rapidly evolving spatial patterns, test a sequence model or radar-based approach—but only after establishing a strong baseline. Teams building dependable infrastructure can also apply monitoring ideas from Building Predictive Maintenance Systems with AI, especially for data freshness, drift, and alert escalation.

For an India-focused ML project, the strongest grant or pilot proposal will specify the venue, data rights, prediction horizon, evaluation protocol, and measurable operational benefit. AI Grants India supports builders turning such clearly scoped ideas into deployable systems.

Apply for AI Grants India