Time Series Forecasting Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Time series forecasting models require significant time and resources for training and re-training due to the ever-changing nature of time series data, leading to inefficiencies and increased latency in processing.
Innovation Solution
A forecasting service system that optimizes computing resources by selecting an appropriate compute instance based on the attributes of the time series dataset, allowing for either dedicated or shared use of pre-provisioned instances to train models efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple machine learning models are trained and validated to select the most appropriate model for time series forecasting, then the forecasting accuracy is improved, but the training time and computational resources increase significantly
Solution Approach 1:
The system pre-trains multiple machine learning models offline before they are needed for forecasting. These pre-trained models are stored and can be quickly deployed when forecasting is required, eliminating the need for time-consuming training at the time of use. This allows the system to maintain high forecasting accuracy through multiple model options while avoiding the time penalty of training them on-demand.
Solution Approach 2:
The system creates and maintains multiple copies of trained models (different model architectures and configurations) that have been pre-trained on historical data. When a forecasting request comes in, the system can select from these existing model copies rather than training a new model, thus maintaining accuracy through model diversity while significantly reducing training time.
2Measurement precision
If models are re-trained frequently to adapt to changing time series data, then the model accuracy is maintained, but the computational resources and processing time increase
Solution Approach 1:
The system performs preliminary training of multiple model variants offline in advance. When data changes occur, the system can quickly re-train or swap between pre-trained models rather than performing full re-training, maintaining accuracy while improving processing efficiency.
Solution Approach 2:
The system develops a universal framework that supports multiple model types and training configurations. This multi-functional system can handle different data scenarios with appropriate pre-trained models, reducing the need for frequent re-training while maintaining accuracy across varying conditions.
3Reliability
If dedicated compute instances are provisioned for model training, then the training reliability and performance are improved, but the resource utilization efficiency decreases
Solution Approach 1:
The system merges multiple training workloads onto shared compute instances rather than using dedicated instances for each model. By consolidating training tasks and scheduling them appropriately, the system maintains training reliability through proper resource allocation while improving overall resource utilization efficiency by eliminating idle dedicated instances.
Solution Approach 2:
The system pre-trains models during periods of lower demand and stores the trained models for later use. This allows dedicated compute resources to be used intensively during pre-training phases rather than being continuously allocated, improving resource utilization while maintaining training reliability through scheduled pre-training operations.
Data Source
AI summary
A time series forecasting system is disclosed that receives a forecast request identifying a forecast to be made for a future time point and a time series dataset to be used for making the requested forecast. The system selects a model to be used for making the requested forecast and determines a set of one or more attributes of the time series dataset. Based on the set of attributes, the system selects a compute instance either from a shared node pool or a dedicated instance having a shape to be used for training the selected model and generating a forecast. The system then generates a forecast result for the future time point using the trained model and provides the forecast result as a response to the forecast request.