BiLSTM Temporal Convolution for Multi-Horizon Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time series forecasting methods, such as classical Holt-Winters and ARIMA models, are ineffective in modeling highly nonlinear time series, and LSTM-based approaches fail to explore temporal patterns on long future horizons, leading to inadequate predictions.
Innovation Solution
A bi-directional long-short term memory network (BiLSTM) with temporal convolution and context selection is employed to capture multi-scale temporal contexts, using input feature vectors from future time steps, embedded and concatenated to generate accurate forecasting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If LSTM-based approaches are used for time series forecasting, then nonlinear patterns can be captured, but temporal patterns on long future horizons cannot be explored effectively
Solution Approach 1:
The patent divides the forecasting task into multiple segments by applying temporal convolution at different scales (e.g., short-term, medium-term, long-term windows). Each scale captures temporal patterns at its specific horizon, and the results are aggregated to produce comprehensive multi-horizon forecasts. This segmentation allows the model to effectively explore temporal patterns across different future horizons rather than treating them as a single uniform task.
Solution Approach 2:
The patent introduces a new dimension by applying temporal convolution operations across multiple time scales simultaneously. Instead of using a single LSTM sequence, the model adds a temporal scale dimension by convolving hidden states with different window sizes (e.g., 3, 5, 7, 14 days). This multi-scale temporal dimension enables the model to capture patterns at various forecasting horizons in parallel, resolving the limitation of traditional LSTM approaches that struggle with long-horizon temporal patterns.
2Measurement precision
If multi-scale temporal contexts are captured using temporal convolution, then forecasting accuracy improves, but computational complexity increases
Solution Approach 1:
The computational complexity is managed by segmenting the temporal convolution into distinct scales (e.g., short-term 3-day, medium-term 7-day, long-term 14-day windows). Each scale processes a specific temporal range independently, allowing for optimized computation at each level rather than processing all temporal relationships uniformly. This segmentation reduces the overall computational burden while maintaining high forecasting precision through multi-scale analysis.
Solution Approach 2:
The patent merges the results from multiple temporal scales by aggregating the context features obtained from different convolution windows. Instead of maintaining separate models for each scale, the approach combines the extracted temporal patterns into a unified forecasting framework. This merging reduces model complexity by integrating multiple scales into a single cohesive system while preserving the precision benefits of multi-scale analysis.
Data Source
AI summary
A system and a method for time series forecasting. The method includes: providing input feature vectors corresponding to a plurality of future time steps; performing bi-directional long-short term memory network (BiLSTM) on the input feature vectors to obtain hidden outputs corresponding to the plurality of future time steps; for each future time step: performing temporal convolution on the hidden outputs using a plurality of temporal scales to obtain context features at the plurality of temporal scales, and summating the context features at the plurality of temporal scales using a plurality of weights to obtain multi-scale context features; and converting the multi-scale context features to obtain the time series forecasting corresponding to the future time steps.