US 12,493,771 B2
Deep learning model for energy forecasting
Richa Chauhan, Pune (IN); Harish Yadav, Maharashtra (IN); Hemil Shah, Maharashtra (IN); Kanchan Kamat, Wanowrie (IN); Arnulfo D. de Castro, Durham, NC (US); and Tae Yoon Lee, Johns Creek, GA (US)
Assigned to SAS Institute, Inc., Cary, NC (US)
Filed by SAS Institute Inc., Cary, NC (US)
Filed on Jan. 11, 2024, as Appl. No. 18/410,742.
Claims priority of provisional application 63/536,030, filed on Aug. 31, 2023.
Claims priority of application No. 202311031512 (IN), filed on May 3, 2023.
Prior Publication US 2024/0370697 A1, Nov. 7, 2024
Int. Cl. G06N 3/045 (2023.01); G06N 20/20 (2019.01)
CPC G06N 3/045 (2023.01) [G06N 20/20 (2019.01)] 30 Claims
OG exemplary drawing
 
1. A system comprising:
one or more processors; and
one or more memories including program code that is executable by the one or more processors to:
receive an input from a user indicating a target variable to be forecasted over a future time window, wherein the target variable relates to energy consumption or energy production;
determine a plurality of independent variables that influence the target variable, wherein the plurality of independent variables includes a plurality of temporal variables and a plurality of environmental variables;
generate a set of candidate variables for input to a random forest classifier, wherein the set of candidate variables includes the plurality of independent variables and additional variables, each of the additional variables being a new variable defined using a mathematical expression that includes at least two of the plurality of independent variables serving as operands in the mathematical expression;
execute the random forest classifier that is configured to identify, from the set of candidate variables, a subset of candidate variables having at least a threshold level of influence on the target variable, wherein the random forest classifier is configured to output the identified subset of candidate variables;
automatically construct an architecture of a machine-learning model based on a set of hyperparameter values and the identified subset of candidate variables, wherein the architecture of the machine-learning model is automatically constructed to receive the identified subset of candidate variables as inputs and generate a forecast of the target variable as an output, the architecture of the machine-learning model including at least:
an input layer configured to receive the identified subset of candidate variables;
a reshape layer coupled to the input layer;
one or more convolution layers coupled to the reshape layer;
one or more LSTM layers coupled to the one or more convolution layers;
one or more fully connected layers coupled to the one or more LSTM layers; and
an output layer coupled to the one or more fully connected layers;
after automatically constructing the machine-learning model, train the machine-learning model using historical data that indicates previous values of the target variable over a prior time window;
after training the machine-learning model, execute the machine-learning model to generate the forecast indicating future values for the target variable over the future time window; and
transmit an output to the user indicating the forecast.