Network demand forecasting

• US 10,650,396 B2
• Filed: 06/05/2018
• Issued: 05/12/2020
• Est. Priority Date: 04/04/2014
• Status: Active Grant

• Current Assignee: Kyndryl Incorporated (Kyndryl Holdings, Inc.)
• Original Assignee: International Business Machines Corporation
• Inventors: Baughman, Aaron K., Kozloski, James R., O'Connell, Brian M.
• Primary Examiner(s): Sterrett, Jonathan G
• Application Number:
  US16/000,183
• Publication Number:
  US 20180285903A1
• Time in Patent Office: 707 Days
• CPC Class Codes: G06Q 10/04 Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" market predictions or forecasting for commercial activities G06Q30/0202
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: H04L 41/145 involving simulating, designing, planning or modelling of a network
• CPC Class Codes: H04L 41/147 for predicting network behaviour
• CPC Class Codes: H04L 41/149 for prediction of maintenance

Abstract

Approaches for predicting demand spikes and cyclic demand for web-related content is provided. An approach includes generating multiple models for predicting demand. The approach further includes selecting a model from the multiple models. The approach further includes inputting information into the model. The approach further includes generating, by at least one computing device, a predicted demand spike using the model and the information. The approach further includes generating a total predicted demand curve by combining the predicted demand spike with predicted cyclical demand.

Claims

1. A method for predicting demand of networked computing resources, comprising:

selecting a reference model with a lowest level of error among multiple models for predicting website data demand of an event;

training the selected reference model by boosting the selected reference model;

training the selected reference model by bagging the selected reference model;

selecting one of the boosted selected reference model, the bagged selected reference model, and the selected reference model with the lowest level of error;

generating, by at least one computing device, a predicted demand spike curve using the selected one of the boosted selected reference model, the bagged selected reference model, and the selected reference model and historical information which corresponds to data demand of events similar to the event;

evaluating an accuracy of the selected one of the boosted selected reference model, the bagged selected reference model, and the selected reference model based on a dynamically updated total predicted demand curve by determining a success level of the selected one of the boosted selected reference model, the bagged selected reference model, and the selected reference model for a specified period of time during the event;

switching to another model for a next period of time during the event in response to the selected one of the boosted selected reference model, the bagged selected reference model, and the selected reference model not achieving the success level for the specified period of time during the event; and

provisioning based on the evaluating and the switching, by the at least one computing device, a plurality of website servers based on an updated demand curve.

2. The method of claim 1, wherein the selecting the model includes:

inputting training data into each of the multiple models and comparing a level of error for each of the multiple models.

3. The method of claim 1, further comprising inputting online information.

4. The method of claim 3, wherein the predicted demand spike is associated with a sporting event.

5. The method of claim 4, wherein the online information is associated with current demands for content of a website and information retrieved from social networking websites.

6. The method of claim 1, further comprising updating the total predicted demand curve by inputting the historical information and online information.

7. The method of claim 1, further comprising:

generating an additional model, derived from the selected reference model, such that the additional model includes a weighted variable for time; and

analyzing the additional model with the weighted variable for time associated with the additional model to determine whether the additional model generates outputs that meet threshold values.

8. The method of claim 7, further comprising:

replacing the additional model with another additional model when the outputs meet the threshold values; and

continuing to use the additional model when the outputs do not meet the threshold values.

9. The method of claim 1, wherein the predicted demand spike provides a spike demand value.

10. The method of claim 9, wherein the spike demand value is added to a value associated with the cyclical demand curve.

11. The method of claim 1, wherein the event comprises a web exclusive event featuring a player associated with a sporting event.

12. The method of claim 1, wherein the total predicted demand curve is generated by averaging values of the predicted demand spike curve with values of the predicted cyclical demand.

13. The method of claim 1, further comprising:

generating a total predicted demand curve by combining the predicted demand spike curve with predicted cyclical demand; and

dynamically updating the total predicted demand curve by using updated historical information corresponding to a sliding window of time.

14. The method of claim 1, wherein the success level is determined by generating a graph.