LSTM-based Flow Prediction

Hongzhi Wang, Yang Song, Shihan Tang

Abstract

Predicting production flow in high-speed industrial environments is critical for logistics and operational efficiency. Traditional time-series forecasting models often struggle with the high dimensionality and complex periodicities inherent in industrial sensor data. This paper presents an optimized Long Short-Term Memory (LSTM) framework that integrates multivariate tuning and periodic analysis. By utilizing Random Forest for feature selection and a custom tuning module to identify optimal time-step intervals, the proposed method achieves a 54.05% improvement in prediction accuracy over standard LSTM models on real-world industrial boiler data.

Introduction

Industrial production lines generate vast amounts of high-frequency data from diverse sensors monitoring pressure, temperature, and flow rates. Accurate forecasting of these flows is essential for optimizing supply chains and preventing costly downtime. However, standard deep learning approaches face challenges such as high dimensionality, long-term dependencies spanning thousands of data points, and hidden periodic cycles. Researchers from the Harbin Institute of Technology address these issues by proposing a refined LSTM-based architecture tailored for industrial contexts.

Problem Statement

Standard predictive models encounter three primary obstacles when applied to factory floor data:

  1. High Dimensionality: Correlated sensors create noise that can mislead models.
  2. Temporal Dependencies: High sampling rates require models to maintain memory over extremely long sequences.
  3. Opaque Periodicity: Physical processes often follow cycles that are not immediately obvious, making the selection of appropriate time-series windows difficult.

Proposed Methodology

The proposed system architecture consists of three core modules designed to preprocess, model, and tune industrial time-series data.

System architecture showing the interaction between the Tuning Module, LSTM units, and Data Conversion Module.

Data Conversion Module

To mitigate the "curse of dimensionality," a Random Forest regressor is employed to rank feature importance. The module retains only the top variables contributing to 95% of the cumulative correlation, effectively pruning irrelevant sensor noise.

LSTM Modeling Module

The core predictive engine is a three-layer LSTM network consisting of Input, Hidden, and Output layers. LSTMs are specifically chosen for their capacity to handle nonlinearities and store information over long sequences through gated mechanisms.

Tuning Module

The innovation lies in the Tuning Module, which identifies the natural periodicity of the production flow. By analyzing zero-crossing points in normalized data, the system identifies significant periodic intervals. It then uses a minimum heap to select the five most likely periods and iterates through them to find the one that minimizes the Root Mean Square Error (RMSE). This process reshapes the raw time series into supervised learning sequences that align with the physical process's rhythm.

Visualization of the data transformation process before and after reshaping based on identified periodicity.

Experimental Setup

The model was evaluated using a real-world Industrial Boiler Dataset from a power plant in Harbin, China. The dataset comprises over 400,000 data points across 70 dimensions. A holdout method was used, with 2/3 of the data for training and 1/3 for testing.

Results and Discussion

The optimized LSTM was compared against several baseline models, including CNNs and traditional LSTMs. The results, summarized in the table below, demonstrate the superior performance of the proposed approach.

Algorithm RMSE
Random Forest 40.21
CNN 21.97
Backpropagation (BP) 20.90
Traditional LSTM 19.87
Optimized LSTM (Proposed) 9.13

The optimized model achieved a 54.05% reduction in RMSE compared to the traditional LSTM. The rapid convergence of loss curves and consistency between training and testing sets suggest high robustness and minimal overfitting.

Conclusion

By explicitly incorporating periodicity and feature pruning, the proposed LSTM framework significantly enhances flow prediction in industrial settings. While the current heuristic for period selection is effective, future work may explore more generalized tuning methods and evaluate the model's performance across different industrial sectors such as chemical manufacturing.

Reference: Wang, H., Song, Y., & Tang, S. (2019). LSTM-based Flow Prediction. arXiv:1908.03571v1 [cs.LG].

Summary updated

BibTeX
@misc{wang2019lstm,
  author = {Hongzhi Wang and Yang Song and Shihan Tang},
  title = {{LSTM-based Flow Prediction}},
  year = {2019},
  eprint = {1908.03571},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.03571},
}
Figure