Time-Series Resource Prediction Using Error-Model Selection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing time series forecasting methods often result in poor performance due to the use of a single prediction method, which may not be optimal for all time series, and combining multiple methods increases computational expense and risk of errors.

Innovation Solution

A method that extracts feature representations from time series data, uses trained error models to evaluate prediction errors for multiple models, selects a subset of models with the lowest expected errors, and applies these models to predict future resource usage, optimizing the prediction process while reducing complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If multiple prediction methods are used to improve prediction accuracy, then prediction reliability is improved, but computational expense increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational expense
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system pre-computes and stores error models for multiple prediction methods before actual prediction is needed. During prediction, the system queries these pre-trained error models to estimate expected errors for different methods, avoiding the need to actually execute all prediction methods. This preliminary preparation enables the system to select the best prediction method without performing all possible predictions, thus reducing computational expense while maintaining high prediction accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system introduces error models as intermediary components that mediate between the prediction methods and the final prediction output. Instead of directly comparing and executing multiple prediction methods, the error models serve as intermediaries that provide estimated error metrics, allowing the system to indirectly evaluate and select the best prediction method. This intermediary approach enables accurate prediction method selection without the computational cost of executing all prediction methods.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If multiple prediction methods are used to handle different time series, then adaptability is improved, but device complexity increases

Engineering Contradiction:
Improvemethod diversityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system segments the prediction process into independent components: multiple prediction methods are maintained as separate modules, each with its own pre-trained error model. This segmentation allows the system to manage complexity by organizing prediction methods as discrete, independently manageable units. When making a prediction, the system selectively activates only the relevant segments (prediction methods) based on the time series characteristics, rather than managing all methods as a monolithic complex system.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system implements dynamic selection of prediction methods based on real-time error estimation. Instead of using a fixed set of prediction methods, the system dynamically determines which methods to apply by querying error models and selecting the method with the lowest expected error. This dynamic approach allows the system to adapt to different time series patterns while managing complexity through on-demand method selection rather than maintaining all methods as static, always-active components.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4357975B1Improved prediction methods of resource usage
Publication Date: 2025.12.31 ORANGE SA
  • EP4357975B1 patent drawingFigure 1
  • EP4357975B1 patent drawingFigure 2
  • EP4357975B1 patent drawingFigure 3

AI summary

The invention relates to a computer-implemented method comprising: extracting a feature representation of a time series of observations of resource usage or at least one physical quantity ; for each prediction model in a set of prediction models: using a trained error model associated with the resource prediction model to obtain an expected prediction error of the resource prediction model for said feature representation ; selecting a subset of said set of prediction models based at least on the expected prediction errors ; using said subset of prediction models to predict a future resource usage or at least one physical quantity from said time series of observations.