Multi-Step Time-Series Forecasting via Residual Error Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional time-series prediction methods using recurrent neural networks (RNNs) are limited to handling single long data sequences, failing to efficiently manage multiple data sequences with varying lengths and minimal changes, which is common in retail industry sales forecasting where thousands of products with different price rates are involved.
Innovation Solution
A processor-implemented method for multiple time-series multi-step data forecasting that forms unscaled and scaled subsequence pairs, builds encoder-decoder networks, and iteratively refines forecasting models to handle multiple data sequences by classifying them into successful and unsuccessful sets based on error metrics, and constructs residual error forecasting models for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single forecasting model is used for all data sequences, then the device complexity is reduced, but the prediction accuracy deteriorates for sequences with varying characteristics
Solution Approach 1:
The patent segments the single forecasting model into multiple specialized models, where each model is trained on a specific subset of data sequences with similar characteristics (e.g., similar variance, seasonality patterns, or domain features). This segmentation allows each model to specialize in capturing the unique patterns of its assigned data sequences, thereby improving prediction accuracy without requiring a single overly complex universal model.
Solution Approach 2:
The patent implements a dynamic model selection mechanism that adapts the choice of forecasting model based on the characteristics of the input data sequence. The system dynamically determines which specialized model to use by analyzing features of the incoming data, such as variance levels, seasonal patterns, or other statistical properties, and selects the most appropriate pre-trained model for that specific data type, thus optimizing accuracy for varying data characteristics.
2Adaptability or versatility
If conventional RNN techniques are applied to multiple data sequences, then the adaptability to different data types is improved, but the handling efficiency of short and varying sequences deteriorates
Solution Approach 1:
The patent applies preliminary actions by pre-processing and normalizing each data sequence before feeding it to the forecasting model. This includes handling short sequences through data augmentation techniques, filling missing values, and normalizing varying scales. By performing these preparatory steps in advance, the system ensures that the forecasting model receives standardized input, improving both adaptability to different data types and processing efficiency.
Solution Approach 2:
The patent dynamically adjusts model parameters based on the characteristics of each data sequence. For short sequences, the system modifies hyperparameters such as sequence length, learning rate, or model depth to suit the limited data available. For sequences with varying variance or seasonality, the system adjusts parameters like window size, aggregation levels, or seasonal decomposition parameters, thereby optimizing forecasting efficiency for each specific data type without requiring a completely different model architecture.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
This disclosure relates generally to methods and systems for multiple time-series data forecasting using recurrent neural networks (RNNs). Conventional techniques in the art for the time-series prediction are limited to deal with one long data sequence and a single forecasting model may not be sufficient and efficient to cover the multiple short data sequences. The present disclosure makes use of greedy recursive procedure to build a set of multi-step forecasting models that covers the multiple data sequences, using the recurrent neural network (RNN) models. The one or more multi-step residual error forecasting models makes the forecasting resulting from the set of multi-step forecasting models, accurate and efficient. The set of multi-step forecasting models are useful for various forecasting applications such as prediction of the sales for retail industries, prediction of power consumption for households, the prediction of traffic occupancy across roads, and so on.