Neural Network Initialization
AssigneePassiveLogic, Inc.
InventorsTroy Aaron Harvey, Jeremy David Fillingim
Life of this patent
Estimating the term…
Abstract
A neural network representing a controlled space can be initialized by collecting state time series data that affects the controlled space such as weather, and also collecting sensor data from the controlled space at the same time. The time series data is used as input to a neural network that models the controlled space until an area in the neural network equivalent to the sensor is at or near the sensor state at a given time.
Claims20
A computer-implemented method for reducing compute time required to operate a thermodynamic model of a controlled space, the method comprising:
- receiving, by one or more processors, time-series state data representing a physical state that affects the controlled space during a historical time period ending at a reference time;
- receiving, by the one or more processors, sensor data from a sensor located in the controlled space, the sensor data representing a measured state of a portion of the controlled space at the reference time;
- storing, in memory, a thermodynamic neural network representing the controlled space, the thermodynamic neural network comprising a plurality of neurons arranged to correspond to physical portions of the controlled space;
- and
- before running the thermodynamic neural network to determine a control operation for equipment of the controlled space, performing a warmup process to establish initialized transient state values for the thermodynamic neural network, wherein the initialized transient state values represent internal thermodynamic states, the warmup process comprising:
- selecting a first portion of the time-series state data;
- propagating the first portion of the time-series state data through the thermodynamic neural network to update transient state values of the plurality of neurons;
- determining whether an updated transient state value of a sensor-corresponding neuron satisfies a stopping condition relative to the measured state represented by the sensor data, the sensor-corresponding neuron corresponding to the portion of the controlled space in which the sensor is located;
- when the stopping condition is not satisfied, selecting a second portion of the time-series state data having a different historical duration than the first portion and propagating the second portion of the time-series state data through the thermodynamic neural network;
- and
- when the stopping condition is satisfied, storing current transient state values of the plurality of neurons as initialized transient state values of the thermodynamic neural network, and using the initialized transient state values as starting values for a later run of the thermodynamic neural network, thereby reducing a number of iterations or processor cycles required for the later run to reach a modeled state corresponding to the controlled space.
A system for reducing compute time required to operate a thermodynamic model of a controlled space, the system comprising:
- one or more processors;
- memory in operational communication with the one or more processors;
- and
- instructions stored in the memory that, when executed by the one or more processors, cause the one or more processors to:
- receive time-series state data representing a physical state that affects the controlled space during a historical time period ending at a reference time;
- receive sensor data from a sensor located in the controlled space, the sensor data representing a measured state of a portion of the controlled space at the reference time;
- and
- store, in the memory, a thermodynamic neural network representing the controlled space, the thermodynamic neural network comprising a plurality of neurons arranged to correspond to physical portions of the controlled space;
- before running the thermodynamic neural network to determine a control operation for equipment of the controlled space, perform a warmup process to establish initialized transient state values for the thermodynamic neural network, wherein the initialized transient state values represent internal thermodynamic states, the warmup process comprising:
- select a first portion of the time-series state data;
- propagate the first portion of the time-series state data through the thermodynamic neural network to update transient state values of the plurality of neurons;
- determine whether an updated transient state value of a sensor-corresponding neuron satisfies a stopping condition relative to the measured state represented by the sensor data, the sensor-corresponding neuron corresponding to the portion of the controlled space in which the sensor is located;
- when the stopping condition is not satisfied, select a second portion of the time-series state data having a different historical duration than the first portion and propagate the second portion of the time-series state data through the thermodynamic neural network;
- and
- when the stopping condition is satisfied, store current transient state values of the plurality of neurons as initialized transient state values of the thermodynamic neural network, and use the initialized transient state values as starting values for a later run of the thermodynamic neural network, thereby reducing a number of iterations or processor cycles required for the later run to reach a modeled state corresponding to the controlled space.
A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations for reducing compute time required to operate a thermodynamic model of a controlled space, the operations comprising:
- receiving time-series state data representing a physical state that affects the controlled space during a historical time period ending at a reference time;
- receiving sensor data from a sensor located in the controlled space, the sensor data representing a measured state of a portion of the controlled space at the reference time;
- storing, in memory, a thermodynamic neural network representing the controlled space, the thermodynamic neural network comprising a plurality of neurons arranged to correspond to physical portions of the controlled space;
- before running the thermodynamic neural network to determine a control operation for equipment of the controlled space, performing a warmup process to establish initialized transient state values for the thermodynamic neural network, wherein the initialized transient state values represent internal thermodynamic states, the warmup process comprising:
- selecting a first portion of the time-series state data;
- propagating the first portion of the time-series state data through the thermodynamic neural network to update transient state values of the plurality of neurons;
- determining whether an updated transient state value of a sensor-corresponding neuron satisfies a stopping condition relative to the measured state represented by the sensor data, the sensor-corresponding neuron corresponding to the portion of the controlled space in which the sensor is located;
- when the stopping condition is not satisfied, selecting a second portion of the time-series state data having a different historical duration than the first portion and propagating the second portion of the time-series state data through the thermodynamic neural network;
- and
- when the stopping condition is satisfied, storing current transient state values of the plurality of neurons as initialized transient state values of the thermodynamic neural network, and using the initialized transient state values as starting values, rather than default starting values, for a later run of the thermodynamic neural network, thereby reducing a number of iterations or processor cycles required for the later run to reach a modeled state corresponding to the controlled space.
Dependencies are inferred from the claim text.