Short-term load forecasting

Patent metadata

Field	Value
Publication number	US-11847591-B2
Application number	US-202016953586-A
Country	US
Kind code	B2
Filing date	Nov 20, 2020
Priority date	Jul 6, 2020
Publication date	Dec 19, 2023
Grant date	Dec 19, 2023

Abstract

A method, computer program, and computer system are provided for load forecasting. Datasets corresponding to source machine learning models and a target domain base model are identified. A set of forecasting models corresponding to the identified datasets are learned. An ensemble model is determined from the learned set of forecasting models based on gradient boosting. An available resource is allocated based on the ensemble model.

First claim

What is claimed is:

1. A method of load forecasting, executable by a processor, comprising:
identifying datasets corresponding to source machine learning models and a target domain base model;
learning a set of forecasting models corresponding to the identified datasets;
determining an ensemble model from the learned set of forecasting models based on gradient boosting; and allocating an available resource based on the ensemble model,
wherein determining the ensemble model comprises:
initializing a residual value,
calculating, for source domains associated with the source machine learning models, a temporary weight value and a loss value;
training a base learner based on the calculated temporary weight value and the loss value; and
updating the residual value based on the trained base learner,
wherein a learned model multiplied by a small step size is added to the ensemble model, and
wherein the forecasting models are learned based on a direction of a negative gradient of a loss function associated with each of the identified datasets, and the ensemble model is identified based on the residual value.

2. The method of claim 1 , wherein the available resource corresponds to an electric load.

3. The method of claim 2 , wherein the datasets correspond to one or more from among a lagged electric load data, historical temperature data, and weekday and weekend data.

4. The method of claim 2 , further comprising:
forecasting a load for a single-family dwelling from among the available electric load; and
allocating electric power corresponding to the forecast load based on the ensemble model.

5. The method of claim 1 , wherein the base learner corresponds to one from among a long-short term memory model learned from the source domains and a linear model learned from a target domain associated with the target domain base model.

6. The method of claim 1 , wherein the temporary weight value re-scales the loss value based on the residual value.

7. The method of claim 1 , further comprising training the ensemble model based on minimizing a negative transfer, wherein the negative transfer corresponds to source domains associated with the source machine learning models and a target domain associated with the target domain base model having a low correlation factor.

8. The method of claim 1 , wherein the ensemble model is determined by a neural network comprising one or more long short-term memory layers.

9. The method of claim 1 , wherein the ensemble model is determined by a linear regression model trained on a target domain.

10. The method of claim 1 , wherein the ensemble model is determined by a linear regression model trained on a target domain.

11. The method of claim 1 , wherein the forecasting models are deep learning models learned from source domains while tackling potential negative transfer, and
wherein at a beginning of a learning process, there are LSTM based forecasting models F 1 . . . F S , trained on the source domains and each boosting iteration t∈{1, . . . , T} performed by the LSTM models trained on source domains and a linear regression model trained on target domain, and
wherein the base learner is identified based on the residual value and added to the ensemble model with a small weight of ε.

12. The method of claim 11 , wherein the base learner is either be a LSTM model (h*) learned from one source domain or a linear model (f*) learned from the target domain as target domain base model, and
wherein h* is chosen among F 1 . . . F S according to its performance on the target domain.

13. A computer system for load forecasting, the computer system comprising:
one or more computer-readable non-transitory storage media configured to store computer program code; and
one or more computer processors configured to access said computer program code and operate as instructed by said computer program code, said computer program code including:
identifying code configured to cause the one or more computer processors to identify datasets corresponding to source machine learning models and a target domain base model;
learning code configured to cause the one or more computer processors to learn a set of forecasting models corresponding to the identified datasets;
determining code configured to cause the one or more computer processors to determine an ensemble model from the learned set of forecasting models based on gradient boosting; and
allocating code configured to cause the one or more computer processors to allocate an available resource based on the ensemble model,

wherein determining the ensemble model comprises:
initializing a residual value,
calculating, for source domains associated with the source machine learning models, a temporary weight value and a loss value;
training a base learner based on the calculated temporary weight value and the loss value; and
updating the residual value based on the trained base learner,
wherein a learned model multiplied by a small step size is added to the ensemble model, and
wherein the forecasting models are learned based on a direction of a negative gradient of a loss function associated with each of the identified datasets, and the ensemble model is identified based on the residual value.

14. The computer system of claim 13 , wherein the available resource corresponds to an electric load.

15. The computer system of claim 14 , wherein the datasets correspond to one or more from among a lagged electric load data, historical temperature data, and weekday and weekend data.

16. The computer system of claim 14 , further comprising:
forecasting a load for a single-family dwelling from among the available electric load; and
allocating electric power corresponding to the forecast load based on the ensemble model.

17. The computer system of claim 13 , wherein the base learner corresponds to one from among a long-short term memory model learned from the source domains and a linear model learned from a target domain associated with the target domain base model.

18. The computer system of claim 13 , further comprising training the ensemble model based on minimizing a negative transfer, wherein the negative transfer corresponds to source domains associated with the source machine learning models and a target domain associated with the target domain base model having a low correlation factor.

19. The computer system of claim 13 , wherein the ensemble model is determined by a neural network comprising one or more long short-term memory layers.

20. A non-transitory computer readable medium having stored thereon a computer program for load forecasting, the computer program configured to cause one or more computer processors to:
identify datasets corresponding to source machine learning models and a target domain base model;
learn a set of forecasting models corresponding to the identified datasets;
determine an ensemble model from the learned set of forecasting models based on gradient boosting; and
allocate an available resource based on the ensemble model,
wherein the computer program is further configured to cause the one or more computer processors to determine the ensemble model by:
initializing a residual value,
calculating, for source domains associated with the source machine learning models, a temporary weight value and a loss value;
training a base learner based on the calculated temporary weight value and the loss value; and
updating the residual value based on the trained base learner,
wherein a learned model multiplied b

Assignees

• Samsung Electronics Co Ltd

Inventors

• Wu Di
• Xu yi tian
• Chen Xi
• Wang Ju
• Jenkin Michael
• Li Hang
• Dudek Gregory Lewis
• Liu Xue

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06Q10/04 Primary

  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 title

• G06N20/20

  Ensemble learning · CPC title