TCAN TensorFlow

TensorFlow implementation of multivariate time series forecasting model introduced in Lin, Y., Koprinska, I., and Rana, M. (2021). Temporal Convolutional Attention Neural Networks for Time Series Forecasting. In 2021 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE.

Dependencies

pandas==1.5.2
numpy==1.23.5
tensorflow==2.11.0
tensorflow_probability==0.19.0
tensorflow_addons==0.19.0
plotly==5.11.0
kaleido==0.2.1

Usage

import numpy as np

from tcan_tensorflow.model import TCAN
from tcan_tensorflow.plots import plot

# Generate some time series
N = 500
t = np.linspace(0, 1, N)
e = np.random.multivariate_normal(mean=np.zeros(3), cov=np.eye(3), size=N)
a = 10 + 10 * t + 10 * np.cos(2 * np.pi * (10 * t - 0.5)) + 1 * e[:, 0]
b = 20 + 20 * t + 20 * np.cos(2 * np.pi * (20 * t - 0.5)) + 2 * e[:, 1]
c = 30 + 30 * t + 30 * np.cos(2 * np.pi * (30 * t - 0.5)) + 3 * e[:, 2]
y = np.hstack([a.reshape(-1, 1), b.reshape(-1, 1), c.reshape(-1, 1)])

# Fit the model
model = TCAN(
    y=y,
    x=None,
    forecast_period=100,
    lookback_period=100,
    quantiles=[0.001, 0.1, 0.5, 0.9, 0.999],
    filters=32,
    kernel_size=7,
    dilation_rates=[1, 2, 4],
    dropout=0,
    alpha=1.5
)

model.fit(
    regularization=0.5,
    learning_rate=0.001,
    batch_size=32,
    epochs=200,
    verbose=1
)

# Generate the forecasts
df = model.forecast(y=y, x=None)

# Plot the forecasts
fig = plot(df=df, quantiles=[0.001, 0.1, 0.5, 0.9, 0.999])
fig.write_image('results.png', scale=4, height=900, width=700)