unifyair/handover_prediction

Mobility Prediction Model

A deep learning-based system for predicting mobility patterns and handover requirements in wireless networks using LSTM neural networks.

Overview

This project implements a sophisticated mobility prediction model that uses historical mobility data to predict when a handover between network cells will be needed. The model leverages LSTM (Long Short-Term Memory) networks to capture temporal patterns in user mobility and network conditions.

Features

• LSTM-based deep learning model for sequence prediction
• Comprehensive feature engineering including:
  • Spatial features (x, y coordinates)
  • Temporal features (velocity, heading)
  • Network metrics (signal strength, SINR, network load, throughput)
  • Time-based cyclical features (hour of day, day of week)
  • Categorical features (pattern type, device type)
• Advanced model architecture with:
  • Dual LSTM layers with dropout for regularization
  • Dense layers for final prediction
  • Binary classification output
• Robust data preparation pipeline
• Early stopping and learning rate reduction callbacks
• Comprehensive model evaluation metrics

Model Architecture

The model consists of:

• Input LSTM layer (64 units)
• Dropout layer (0.3)
• Second LSTM layer (32 units)
• Dropout layer (0.3)
• Dense layer (32 units, ReLU activation)
• Output layer (1 unit, Sigmoid activation)

Performance Metrics

The model is evaluated using:

• Accuracy
• Precision
• Recall
• AUC (Area Under the Curve)

Data Requirements

The model expects the following features:

• Spatial data: x, y coordinates
• Mobility metrics: velocity, heading
• Network metrics: signal_strength, sinr, network_load, throughput_mbps
• Temporal data: timestamp
• Categorical data: pattern_type, device_type (optional)
• Target variable: handover_needed

Usage

# Prepare your data
X, y, scaler, feature_names = prepare_lstm_data_robust(
    data,
    sequence_length=20,
    prediction_horizon=5
)

# Build and train the model
model = build_mobility_prediction_model(input_shape=(X.shape[1], X.shape[2]))
model.fit(X_train, y_train, validation_data=(X_val, y_val))

Dependencies

• TensorFlow
• NumPy
• Pandas
• Scikit-learn

Model Training

The model includes several training optimizations:

• Early stopping with patience=10
• Learning rate reduction on plateau
• Batch size of 32
• Adam optimizer with learning rate 0.001
• Binary cross-entropy loss function

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.