Localized Temporal Forecasting Model for Event-Driven Time Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Temporal forecasting models often fail to accurately predict near-term system parameters, especially when event data is involved, as they do not adequately account for intermittent or asynchronous events, leading to inaccurate forecasts.
Innovation Solution
A localized temporal forecasting model is trained using filtered historical time series data arrays that are similar to recent data arrays, incorporating event data to improve forecast accuracy by selecting and filtering historical data based on event similarities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all available historical data is used to train the forecasting model, then the model has sufficient training data, but the forecast accuracy for recent data patterns deteriorates
Solution Approach 1:
The patent segments the historical time series data into multiple groups based on similarity to recent data patterns. Instead of treating all historical data uniformly, the system divides it into relevant and irrelevant segments, training the model only on segments that are similar to recent patterns. This segmentation resolves the contradiction by providing sufficient training data from relevant historical periods while excluding dissimilar data that would degrade forecast accuracy.
Solution Approach 2:
The patent applies local quality by selecting specific subsets of historical data that have local similarities to recent data patterns. Rather than using global historical data, the system identifies and uses only those historical periods that locally match current conditions. This ensures the training data has the appropriate quality and characteristics for accurate forecasting of recent patterns.
2Device complexity
If traditional temporal forecasting models are used, then the model structure is simple, but the ability to account for event data deteriorates
Solution Approach 1:
The patent merges traditional temporal forecasting with event data analysis by integrating event detection and filtering capabilities into the forecasting pipeline. The system combines time series data with event data, using event information to filter and select relevant historical periods for training. This merging allows the model to maintain a relatively simple structure while significantly improving reliability in the presence of events.
Solution Approach 2:
The patent introduces event data as an intermediary element that mediates between the historical time series data and the forecasting model. Events serve as filtering criteria that select which historical periods are relevant for training. This intermediary mechanism allows the model to account for events without requiring complex event-driven architecture, maintaining simplicity while improving reliability.
3Quantity of substance
If historical data far from recent patterns is used for training, then the training dataset is large, but the relevance to current conditions deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-filtering and selecting historical time series data based on similarity to recent patterns before training the model. The system proactively identifies and excludes historical periods that are dissimilar to current conditions, ensuring that only relevant data is used for training. This preliminary selection maintains an adequate training dataset size while ensuring high relevance to current conditions.
Solution Approach 2:
The patent applies parameter changes by dynamically adjusting the selection criteria for historical data based on the similarity between historical and recent patterns. The system changes the relevant parameters (time periods, conditions) of historical data that are included in the training set, ensuring adaptability to current conditions while maintaining sufficient training data volume.
Data Source
AI summary
Disclosed herein are systems, computer-readable media, and methods related to modeling on multivariate time series data overlaid with event data. In particular, some examples involve selecting one or more historical time series data arrays similar to a recent time series data array and filtering the similar historical time series data arrays based on event data. Some examples can also involve training a localized temporal forecasting model using the filtered historical time series data arrays. Some examples can include building and/or training the localized temporal forecasting model at or near a time that a forecast is needed.