Using TensorFlow.js for Time Series Analysis

Using TensorFlow.js for Time Series Analysis

Time series data involves sequential observations indexed in time order, such as stock prices, weather data, or sales trends. TensorFlow.js provides robust tools for handling time series tasks like forecasting, anomaly detection, and pattern recognition.

Key Steps in Time Series Analysis with TensorFlow.js

1. Data Preprocessing: Transform raw time series data into a format suitable for machine learning models.

2. Model Selection: Use architectures like RNNs, LSTMs, or GRUs for capturing temporal dependencies.

3. Training and Evaluation: Train models on historical data and evaluate their performance.

4. Prediction and Forecasting: Use trained models to predict future values.

Step 1: Preprocessing the Data

Time series data needs to be structured into sequences for input into models.

import * as tf from '@tensorflow/tfjs';

// Example time series data (e.g., daily temperatures)
const rawData = [30, 32, 34, 33, 35, 37, 40, 38, 36, 35];

// Create sequences and labels
const windowSize = 3; // Number of previous steps used for prediction
const sequences = [];
const labels = [];

for (let i = 0; i < rawData.length - windowSize; i++) {
  sequences.push(rawData.slice(i, i + windowSize)); // Input sequence
  labels.push(rawData[i + windowSize]); // Target value
}

// Convert to tensors
const xs = tf.tensor2d(sequences); // Shape: [numSamples, windowSize]
const ys = tf.tensor1d(labels); // Shape: [numSamples]

Step 2: Designing the Model

Recurrent neural networks (RNNs) or LSTMs are well-suited for time series due to their ability to capture sequential dependencies.

const model = tf.sequential();

// Add an LSTM layer
model.add(tf.layers.lstm({
  units: 50,                 // Number of LSTM units
  inputShape: [windowSize, 1], // Sequence length and number of features
  returnSequences: false,    // Output only the final time step
}));

// Add a dense layer for regression output
model.add(tf.layers.dense({ units: 1 })); // Predict a single value

// Compile the model
model.compile({
  optimizer: tf.train.adam(),
  loss: 'meanSquaredError',
});

// Display model summary
model.summary();

Step 3: Training the Model

Train the model on the time series data.

(async () => {
  const reshapedXs = xs.reshape([xs.shape[0], windowSize, 1]); // Reshape for LSTM input
  await model.fit(reshapedXs, ys, {
    epochs: 50,
    batchSize: 16,
    validationSplit: 0.2, // Use 20% of data for validation
    callbacks: {
      onEpochEnd: (epoch, logs) => {
        console.log(`Epoch ${epoch + 1}: Loss = ${logs.loss}`);
      },
    },
  });
})();

Step 4: Making Predictions

Use the trained model to forecast future values.

const testSequence = tf.tensor2d([[37, 40, 38]]).reshape([1, windowSize, 1]); // Example input
const prediction = model.predict(testSequence);
prediction.print(); // Predicted value

Advanced Techniques for Time Series Analysis

• Sliding Window Method: Generate overlapping windows of data to increase the training dataset’s size.

• Feature Engineering: Add additional features like day-of-week, holiday flags, or moving averages.

• Attention Mechanisms: Use attention layers to focus on the most relevant parts of the sequence.

• Multi-Step Forecasting: Predict multiple future values simultaneously by adjusting the model’s output layer.