CeNN-Forecaster: Trainable Cellular Neural Networks for Multivariate Long-Term Time Series Forecasting

Long-horizon multivariate time-series forecasting is challenging due to non-stationary dynamics, cross-variate interactions, and error accumulation over extended rollouts. Attention-based Transformers offer capacity but are computationally heavy, while purely linear/MLP models often miss local dynamical structure. We introduce CeNN-Forecaster, a lightweight hybrid that bridges dynamical-systems theory and deep learning by adapting continuous-time Cellular Neural Networks into a trainable discrete architecture. The core layer unrolls a Forward–Euler residual update with learnable causal Conv1D templates for feedback and control, followed by a compact MLP head for multi-horizon prediction. This yields an explicit dynamical prior that preserves locality and causality, trains end-to-end with standard backpropagation, and promotes stable long-range evolution with a small number of parameters. On the ETT benchmarks (ETTh1, ETTh2, ETTm1, ETTm2) in a multi-target setting with a 256-step context and horizons of 96, 192, 336, and 720, CeNN-Forecaster achieves best performance among compared baselines on ETTh1 across all horizons and on long-horizon ETTm1, while remaining highly competitive on ETTh2 and ETTm2. Additionally, experiments conducted within Microsoft’s BatteryML framework confirm the model’s robustness in battery health forecasting, where it outperformed multiple baselines in Remaining Useful Life (RUL) prediction across diverse datasets. The combination of accuracy, long-horizon stability, and practical efficiency positions CeNNForecaster as a compelling alternative to attention-heavy models for real-world forecasting and resource-constrained deployment.