Neural Network Forecasting Time-Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting models for telecommunications networks face challenges due to their low-capacity learning approaches, requiring multiple models for each time-series, leading to high computational costs and the risk of overfitting, especially in dynamic network environments with many data points.
Innovation Solution
A neural network architecture that separates time-series data into groups and applies them to multiple sub-branches, generating separate forecasting coefficients for each output time point, using an inverse Wavelet transform for improved prediction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a forecasting model is trained for each time-series using low-capacity learning approaches, then the model can fit the historical time-series, but the computational cost and memory requirements become tremendous when forecasting thousands of time-series
Solution Approach 1:
The patent combines multiple individual time-series forecasting models into a single unified neural network model. Instead of training separate models for each time-series, the invention processes multiple time-series simultaneously through one model, sharing computational resources and parameters across all series. This merging approach dramatically reduces the total computational cost and memory requirements while maintaining forecasting accuracy for thousands of time-series points.
2Measurement precision
If low-capacity learning approaches are used with careful model selection by forecasting experts, then the model can be fit to historical data, but the approach does not scale in network settings with large numbers of changing data points
Solution Approach 1:
The patent implements a dynamic neural network architecture that automatically adapts to different time-series patterns and changes in the underlying random processes. The model uses dynamic parameter adjustment and learns from multiple time-series simultaneously, enabling it to scale to network environments with thousands of changing data points. The system dynamically adjusts its forecasting approach based on the characteristics of each time-series without requiring manual model selection or expert intervention for each series.
3Reliability
If LSTM layers are used to fit time-series with specific structural constraints, then the model can capture temporal dependencies, but the model capacity is limited by the imposed structure
Solution Approach 1:
The patent segments the time-series processing into multiple parallel processing streams within the neural network, allowing different parts of the model to handle different aspects of temporal patterns. Instead of forcing all time-series through a single rigid LSTM structure, the invention divides the processing into multiple pathways that can capture diverse temporal dependencies simultaneously, thereby increasing overall model capacity while maintaining the ability to capture temporal relationships.
4Productivity
If Fourier transforms are used in the output layer to fit all future points to the same function, then the model can generate forecasts, but the model capacity is limited by this constraint
Solution Approach 1:
The patent applies local quality by allowing different parts of the output to be generated by different functions rather than forcing all future points to conform to a single Fourier transform. The neural network generates forecasts with varying characteristics for different time points and different time-series, enabling each local region of the forecast to have the appropriate quality and pattern. This approach dramatically increases model capacity while maintaining forecast generation capability.
Data Source
AI summary
Systems, methods, and computer-readable medium for forecasting a time-series are provided. In one implementation, a method is configured to include a step of providing a time-series to a neural network including one or more branches for processing one or more portions of the time-series. In each of the one or more branches, the method includes separating the respective portion of the time-series into individual portions and applying each portion to a respective sub-branch of a plurality of sub-branches of the one or more branches. The method also includes generating forecasting coefficients for each output time point in each of the respective sub-branches and providing a forecast of the time-series based at least on the forecasting coefficients.