CNN-LSTM Temperature Forecasting

Undergraduate final project done for a Bachelor's degree in Operational Research.

Accurate temperature forecasting has practical applications across energy planning, agriculture, and urban infrastructure. This project builds a hybrid CNN-LSTM model to forecast daily temperature in Algiers from multivariate climate data, achieving 1.94°C RMSE on an unseen test set. The model is lightweight (90K parameters), making it suitable for resource-constrained environments.

Key Features:

• Combines CNN for local pattern extraction with LSTM for temporal dependency modeling
• Multivariate approach using 5 climate variables
• 8 years of historical weather data (2016-2024)
• Achieves 1.94°C RMSE on unseen test set

Results

Performance Metrics

Metric	Value
RMSE (Test)	1.94°C
MAE (Test)	~1.26°C
Relative Error	6.3%
Parameters	90,945

Model Comparison

Comparison of RMSE across different CNN-LSTM architectures

Predictions vs Actual Temperature

Time series comparison showing excellent tracking of seasonal patterns

Prediction Accuracy

Scatter plot demonstrating strong correlation between predictions and actual values

Model Architecture

Input Shape: (timesteps, 5 features)
    ↓
Conv1D(256 filters, kernel=3, activation=relu)
    ↓
MaxPooling1D(pool_size=2)
    ↓
LSTM(64 units)
    ↓
Dense(32 units, activation=relu)
    ↓
Dense(1 unit, activation=linear)
    ↓
Output: Temperature (°C)

Summary:

• Total Parameters: 90,945
• Trainable Parameters: 90,945
• Optimizer: Adam
• Loss Function: Mean Squared Error (MSE)

Dataset

Overview

• Location: Algiers, Algeria
• Period: June 2016 - April 2024 (8 years)
• Frequency: Daily observations
• Source: rp5.ru Weather Archives

Features (Input Variables)

Variable	Description	Unit
Temperature	Air temperature	°C
Pressure	Atmospheric pressure	hPa
Humidity	Relative humidity	%
Wind Speed	Average wind speed	m/s
Dew Point	Dew point temperature	°C

Data Split

• Training: 80% (2016-2022)
• Validation: 10% (2023)
• Test: 10% (2024)

Temperature Statistics

Set	Min (°C)	Max (°C)	Range (°C)
Actual	8.44	39.96	31.52
Predicted	7.85	32.25	24.40

Methodology

1. Data Preprocessing

• Normalization using MinMaxScaler (0-1 range)
• Sliding window approach for sequence creation
• Handling missing values and outliers

2. Model Training

• Epochs: 50 with Early Stopping (patience=10)
• Batch Size: 32
• Validation Strategy: Temporal split to prevent data leakage
• Callbacks: ModelCheckpoint, EarlyStopping

3. Evaluation

• RMSE (Root Mean Squared Error)
• MAE (Mean Absolute Error)
• Visual analysis of predictions vs actual values

Strengths

- Excellent Accuracy: RMSE of 1.94°C is comparable to operational weather models
- Captures Seasonality: Successfully tracks seasonal temperature variations
- Lightweight and efficient model Only 90K parameters, suitable for resource-constrained environments
- Multivariate Approach: Leverages multiple climate variables for robust predictions
- No Overfitting: Validation and test performance are consistent

Limitations

- Extreme Temperature Underestimation: Model predicts max 32.25°C vs actual 39.96°C
- Extrapolation Challenge

Why Extreme Temperatures Are Difficult

The test set (2024) contains unprecedented heat records not present in the training data (2016–2022), which explains the underestimation. This is a known limitation of neural networks when extrapolating beyond their training distribution. Proposed mitigations include a weighted loss function to penalize extreme temperature errors more heavily, and synthetic augmentation of extreme samples.

Future Improvements

Potential enhancements to consider:

1. Weighted Loss Function: Penalize errors on extreme temperatures more heavily
2. Data Augmentation: Synthesize additional extreme temperature samples
3. Cyclical Features: Encode seasonal patterns (month, day of year)
4. Multi-step Forecasting: Predict temperature for next 7-14 days
5. Transfer Learning: Adapt model to other geographic locations

---