Systems, methods and devices for vector control of permanent magnet synchronous machines using artificial neural networks
An example method for controlling an AC electrical machine can include providing a PWM converter operably connected between an electrical power source and the AC electrical machine and providing a neural network vector control system operably connected to the PWM converter. The control system can include a current-loop neural network configured to receive a plurality of inputs. The current-loop neural network can be configured to optimize the compensating dq-control voltage. The inputs can be d- and q-axis currents, d- and q-axis error signals, predicted d- and q-axis current signals, and a feedback compensating dq-control voltage. The d- and q-axis error signals can be a difference between the d- and q-axis currents and reference d- and q-axis currents, respectively. The method can further include outputting a compensating dq-control voltage from the current-loop neural network and controlling the PWM converter using the compensating dq-control voltage.
1. A method for controlling an AC electrical machine, comprising:
providing a pulse-width modulated (“PWM”) converter operably connected between an electrical power source and the AC electrical machine;
providing a neural network vector control system operably connected to the PWM converter, the neural network vector control system comprising a current-loop neural network and a speed-loop neural network, wherein at least one of the current-loop neural network and the speed-loop neural network is configured to implement a dynamic programming (“DP”) algorithm;
receiving a plurality of inputs at the current-loop neural network, wherein the plurality of inputs comprise:
a d-axis current, i sd , and a q-axis current, i sq ,
a d-axis error signal and a q-axis error signal, wherein the d-axis error signal comprises a difference between i sd and a reference d-axis current, i sd *, and the q-axis error signal comprises a difference between i sq and a reference q-axis current, i sq *,
an integral of d-axis error signal and an integral of q-axis error signal;
a predicted d-axis current signal and a predicted q-axis current signal; and
a feedback compensating dq-control voltage;
outputting a compensating dq-control voltage from the current-loop neural network, wherein the current-loop neural network is configured to optimize the compensating dq-control voltage based on the plurality of inputs; and
controlling the PWM converter using the compensating dq-control voltage,
wherein at least one of the current-loop neural network and the speed-loop neural network is trained to minimize a cost function of the DP algorithm using a backpropagation through time (“BPTT”) algorithm by unrolling a trajectory of the neural network vector control system from an initial state and training the current-loop neural network or the speed-loop neural network based on the cost function of the DP algorithm and the BPTT algorithm.
2. The method of claim 1 , wherein a predicted d-axis current signal further comprises a difference between i sd and a predicted d-axis current, i sd ′, and a predicted q-axis current signal further comprises a difference between i sq and a predicted q-axis current, i sq ′, wherein i sd ′ and i sq ′ are computed using a current prediction model.
3. The method of claim 2 , wherein the current prediction model is based on i sd , i sq and the compensating dq-control voltage at a previous time step and default parameters for the AC electrical machine.
4. The method of claim 1 , wherein the compensating dq-control voltage is adjusted by a stabilization matrix, the stabilization matrix being based on default parameters for the AC electrical machine.
5. The method of claim 1 , wherein the method further comprises:
receiving a plurality of inputs at the speed-loop neural network, wherein the plurality of inputs comprise:
a speed of the AC electrical machine, ω m ,
a speed error signal comprising a difference between ω m and a reference speed, ω m *,
an integral of speed error signal;
a predicted speed signal; and
a feedback drive torque signal; and
outputting a drive torque signal, τ em , from the speed-loop neural network, wherein the speed-loop neural network is configured to optimize the drive torque signal, τ em , based on the plurality of inputs, and wherein the drive torque signal, τ em , is converted into the reference q-axis current, i sq *.
6. The method of claim 5 , wherein a predicted speed signal further comprises a difference between ω m and a predicted speed signal, ω m ′, wherein ω m ′ is computed using a speed prediction model.
7. The method of claim 5 , wherein the drive torque signal, τ em , is adjusted by a drive-torque stabilization matrix, the drive-torque stabilization matrix being based on default parameters for the AC electrical machine.
8. The method of claim 6 , wherein the speed prediction model is based on ω m and τ em at a previous time step and default parameters for the AC electrical machine.
9. The method of claim 1 , wherein at least one of the current-loop neural network and the speed-loop neural network comprises a multi-layer perceptron including a plurality of input nodes, a plurality of hidden layer nodes and a plurality of output nodes.
10. The method of claim 9 , wherein each of the nodes is configured to implement a hyperbolic tangent function.
11. The method of claim 1 , wherein the AC electrical machine is at least one of a permanent magnet synchronous machine and an induction machine.
12. A system for controlling an AC electrical machine, comprising:
a pulse-width modulated (“PWM”) converter operably connected between an electrical power source and the AC electrical machine;
a neural network vector control system operably connected to the PWM converter, the neural network vector control system comprising a current-loop neural network and a speed-loop neural network, wherein at least one of the current-loop neural network and the speed-loop neural network is configured to implement a dynamic programming (“DP”) algorithm, said current-loop neural network configured to;
receive a plurality of inputs at the current-loop neural network, wherein the plurality of inputs comprise:
a d-axis current, i sd , and a q-axis current, i sq ,
a d-axis error signal and a q-axis error signal, wherein the d-axis error signal comprises a difference between i sd and a reference d-axis current, i sd *, and the q-axis error signal comprises a difference between i sq and a reference q-axis current, i sq *,
an integral of d-axis error signal and an integral of q-axis error signal;
a predicted d-axis current signal and a predicted q-axis current signal; and
a feedback compensating dq-control voltage; and
output a compensating dq-control voltage from the current-loop neural network, wherein the current-loop neural network is configured to optimize the compensating dq-control voltage based on the plurality of inputs, and wherein neural network vector control system controls the PWM converter using the compensating dq-control voltage,
wherein at least one of the current-loop neural network and the speed-loop neural network is trained to minimize a cost function of the DP algorithm using a backpropagation through time (“BPTT”) algorithm by unrolling a trajectory of the neural network vector control system from an initial state and training the current-loop neural network or the speed-loop neural network based on the cost function of the DP algorithm and the BPTT algorithm.
13. The system of claim 12 , wherein a predicted d-axis current signal further comprises a difference between i sd and a predicted d-axis current, i sd ′, and a predicted q-axis current signal further comprises a difference between i sq and a predicted q-axis current, i sq ′, wherein i sd ′ and i sq ′ are computed using a current prediction model.
14. The system of claim 13 , wherein the current prediction model is based on i sd , i sq and the compensating dq-control voltage at a previous time step and default parameters for the AC electrical machine.
15. The system of claim 12 , wherein the compensating dq-control voltage is adjusted by a stabilization matrix, the stabilization matrix being based on default parameters for the AC electrical machine.
16. The system of claim 12 , wherein the speed-loop neural network is configured to:
receive a plurality of inputs at the speed-loop neural network, wherein the plurality of inputs comprise:
a speed of the AC electrical machine, ω m ,
a speed error signal comprising a difference between ω m and a reference speed, ω m *,
an integral of speed error signal;
a predicted speed signal; and
a feedback drive torque signal; and
output a drive torque signal, τ em , from the speed-loop neural network, wherein the speed-loop neural network is configured to optimize the drive torque signal, τ em , based on the plurality of inputs, and wherein the drive torque signal, τ em , is converted into the reference q-axis current, i sq *.
17. The system of claim 16 , wherein a predicted speed signal further comprises a difference between ω m and a predicted speed signal, ω m ′, wherein ω m is computed using a speed prediction model.
18. The system of claim 16 , wherein the drive torque signal, τ em , is adjusted by a drive-torque stabilization matrix, the drive-torque stabilization matrix being based on default parameters for the AC electrical machine.
19. The system of claim 17 , wherein the speed prediction model is based on ω m and τ em at a previous time step and default parameters for the AC electrical machine.
20. The system of claim 12 , wherein at least one of the current-loop neural network and the speed-loop neural network comprises a multi-layer perceptron including a plurality of input nodes, a plurality of hidden layer nodes and a plurality of output nodes.
21. The system of claim 20 , wherein each of the nodes is configured to implement a hyperbolic tangent function.
22. The system of claim 12 , wherein the AC electrical machine is at least one of a permanent magnet synchronous machine and an induction machine.