Conditional noise generative adversarial networks with Siamese neural network for longer time series forecasting

Generative adversarial networks have achieved strong results in computer vision, but their use in time series forecasting remains limited. This paper proposes a conditional noise generative adversarial network with a Siamese neural network as discriminator for long-term forecasting. The method combines the simplicity of a linear model with a generative framework, introducing a triplet margin loss to capture relationships between samples and conditional noise to improve sample generation. Experiments on eight open-source datasets show an average improvement of 8.42 percent, and a 192.8 percent gain for longer-term forecasting, with further improvement on a real-world telecommunications dataset.