Auto-Configured Networks for Multi-Scale Multi-Output Time-Series Forecasting
A new AI framework co-designs data alignment, network architecture, and hyperparameters for industrial time-series forecasting.
A team of researchers has introduced a novel auto-configuration framework designed to tackle the complex challenges of industrial time-series forecasting. The system, detailed in the paper "Auto-Configured Networks for Multi-Scale Multi-Output Time-Series Forecasting," addresses the common scenario of predicting multiple targets from asynchronous, multi-source signals. Current methods often fix key components like data alignment strategies or neural network architecture upfront, making it difficult to systematically co-optimize the entire pipeline—from preprocessing to hyperparameters—under realistic computational constraints.
At its core, the framework operates on two levels. For the model itself, the researchers developed a Multi-Scale Bi-Branch Convolutional Neural Network (MS-BCNN). This architecture features separate branches with short and long convolutional kernels, allowing it to simultaneously capture fine-grained local fluctuations and broader long-term trends critical for accurate multi-output regression. For the optimization process, the team created a hierarchical-conditional mixed configuration space that unifies choices for alignment operators, architectural parameters, and training hyperparameters.
To navigate this vast search space efficiently, the framework employs a Player-based Hybrid Multi-Objective Evolutionary Algorithm (PHMOEA). This algorithm is designed to approximate the optimal trade-off frontier—the Pareto front—between prediction error and model complexity within a limited computational budget. Instead of producing a single model, the system's key output is a set of deployable models, giving engineers flexible choices based on whether they prioritize accuracy or efficiency for a given application.
The framework was validated on hierarchical synthetic benchmarks and a real-world industrial dataset from a sintering process. Experiments demonstrated that it outperforms competitive baseline forecasting methods under the same computational budget, proving its effectiveness in automating the design of robust, production-ready forecasting systems.
- Proposes a unified auto-configuration framework that co-designs data alignment, neural network architecture (MS-BCNN), and training hyperparameters for industrial forecasting.
- Uses a Player-based Hybrid Multi-Objective Evolutionary Algorithm (PHMOEA) to efficiently find a Pareto-optimal set of models balancing error and complexity.
- Outperforms existing baselines on synthetic and real-world sintering data, providing engineers with multiple deployable model choices from a single training run.
Why It Matters
This automates the design of efficient, accurate forecasting models for complex industrial systems, reducing engineering time and improving deployment flexibility.