In-situ thermodynamic model training
Using processes and methods described herein, a digital twin of a physical space can train itself using sensors and other information available from the building. In some embodiments, a system to be controlled comprises a controller that is connected to sensors. This controller also has a thermodynamic model of the system to be controlled within memory associated with the controller. The thermodynamic model has neurons that represent distinct pieces of a controlled space, such as a piece of equipment or a thermodynamically coherent section of a building, such as a window. The neurons represent these distinct pieces of the controlled space using parameter values and equations that model physical behavior of state with reference to the distinct piece of the controlled state. A machine learning process refines the thermodynamic model by modifying the parameter values of the neurons, using sensor data gathered from the system to be controlled as ground truth to be matched by behavior of the thermodynamic model. The thermodynamic model may be warmed up by running the model using state data as input.
1 . A device for in-situ control model training, comprising:
a memory storing a neural network model of a system to be controlled, the neural network model comprising neurons which model thermodynamic coherent sections of the system; at least some of the neurons with multiple parameter values;
a processor in communication with the memory configured to:
retrieve from the memory the neural network model, the neural network model being comprised of neurons, the neurons with activation functions;
retrieve sensor data captured by a controller;
determine a threshold value;
run the neural network model producing computed sensor values by iteratively:
compute a cost function value using a cost function, the computed sensor values and the sensor data,
use the cost function to modify at least one parameter value in at least one neuron, and
run the neural network model producing iterated computed sensor values, until the cost function value is at or below the threshold value; and
wherein at least one of the activation functions of a neuron within the neurons of the neural network model comprises at least two equations that are calculated within the neuron to produce one or more outputs of the neuron.
2 . The device of claim 1 , further comprising backpropagation to take a gradient of the cost function backward through the neural network model.
3 . The device of claim 2 , wherein backpropagation is performed using automatic differentiation.
4 . The device of claim 1 , wherein the neurons comprise input neurons and other neurons, and wherein the activation functions of the other neurons use equations to model physical aspects of individual portions of the system to be controlled.
5 . The device of claim 1 , wherein the system to be controlled comprises an automated building, a process control system, an HVAC system, an energy system, or an irrigation system.
6 . The device of claim 1 , wherein the neural network model is warmed up by being run for a period of time which changes neuron parameter values.
7 . The device of claim 6 , further comprising using optimization to update the at least one parameter value.
8 . The device of claim 1 , further comprising state data affecting the system to be controlled being used as input into the neural network model.
9 . The device of claim 1 , wherein the controller is physically within the system to be controlled.
10 . The device of claim 1 , wherein at least some of the neurons have multiple internal parameters.
11 . The device of claim 1 , wherein the at least one parameter value in the neuron models has a state in a physical location associated with the neural network model.
12 . The device of claim 1 , wherein at least a first neuron models a first physical object.
13 . The device of claim 12 , further comprising a second neuron that models a second physical object, wherein the first physical object outputs to the second physical object, and wherein the first neuron outputs to the second neuron.
14 . The device of claim 13 , wherein the first physical object is a wall.
15 . A method of in-situ neural network training implemented by one or more computers, comprising:
retrieving from a memory a neural network model, the neural network model comprising neurons which model thermodynamic coherent sections of a system; at least some of the neurons with multiple parameter values, the neurons with activation functions;
determining a threshold value;
retrieving sensor data captured by a controller;
running the neural network model producing computed sensor values by iteratively:
computing a cost function value using a cost function, the computed sensor values, and the sensor data,
using the cost function value to modify at least one parameter, and
running the neural network model producing computed sensor values, until the cost function value is at or below the threshold value; and
wherein at least one activation function of a neuron within the neurons comprises at least two equations that are calculated within the neuron to produce one or more outputs of the neuron.
16 . The method of claim 15 , wherein an activation function has multiple parameters whose values are passed between neurons.
17 . The method of claim 15 , wherein the sensor data is a time-state curve.
18 . The method of claim 15 , further comprising running the neural network model for a first period, checking a parameter value, and when the parameter value is not within a range of a threshold value, running the neural network model for a second period.
19 . The method of claim 15 , wherein the computed sensor values is a time-state curve.
20 . A non-transitory computer-readable storage medium configured with executable instructions to perform a method for training a model in-situ, the method comprising:
instructions for retrieving from a memory a neural network model, the neural network model comprising neurons which model thermodynamic coherent sections of a system; at least some of the neurons with multiple parameter values, the neurons with activation functions;
instructions for retrieving sensor data captured by a controller;
instructions for running the neural network model producing computed sensor values by iteratively:
determining a threshold value;
computing a cost function value using a cost function, the computed sensor values and the sensor data,
using the cost function value to modify at least one parameter value, and
running the neural network model producing computed sensor values;
until the cost function value is at or below the threshold value; and
wherein at least one activation function of one a neuron comprises at least two equations that are calculated within the neuron to produce one or more outputs of the neuron.