Method and apparatus for determining battery state-of-charge using neural network architecture

• US 6,064,180 A
• Filed: 10/29/1996
• Issued: 05/16/2000
• Est. Priority Date: 10/29/1996
• Status: Expired due to Fees

• Current Assignee: ADVANCED RESEARCH & TECHNOLOGY INSTITUTE, General Motors Corporation (General Motors Company)
• Original Assignee: General Motors Corporation (General Motors Company), Indiana University Foundation (Indiana University)
• Inventors: Chen, Yaobin, Eberhart, Russell Carley, Brost, Ronald David, Sullivan, Michael Scott
• Primary Examiner(s): Tso, Edward H.
• Application Number:
  US08/740,421
• Time in Patent Office: 1,295 Days
• Field of Search: 320/5, 320/12, 320/30, 320/31, 320/32, 320/35, 320/39, 320/106, 320/110, 320/124, 320/125, 320/127, 320/128, 320/132, 320/135, 320/137, 320/148, 320/149, 320/150
• US Class Current: 320/132
• CPC Class Codes: B60L 58/10 for monitoring or controlling batteries
• CPC Class Codes: G01R 31/3648 comprising digital calculation means, e.g. for performing an algorithm
• CPC Class Codes: G01R 31/367 Software therefor, e.g. for battery testing using modelling or look-up tables
• CPC Class Codes: Y02T 10/70 Energy storage systems for electromobility, e.g. batteries

Abstract

A neural network characterized by a minimal architecture suitable for implementation in conventional microprocessor battery pack monitoring hardware includes linear and non-linear processing elements and battery parameter measurements representative of real time and temporal quantities whereby state of charge estimations actually converge with 100% and 0% states-of-charge.

Claims

1. A neural network system for determining a state-of-charge of an electro-chemical device in a motor vehicle used for providing power to an electric machine in a drive mode and for receiving power from the electric machine in a regenerative mode, said electro-chemical device including a plurality of electro-chemical modules, said neural network system comprising, in combination:

an input layer including a plurality of input layer processing elements;

a hidden layer including a plurality of non-linear hidden layer processing elements;

an output layer including a linear output processing element for providing a scalar output representative of the state-of-charge, said linear output procesing element providing non-asymptotic convergence of the scalar output to battery state-of-charge boundaries;

respective sets of pre-established inter-layer weights having been established in accordance with predetermined input training vectors and corresponding state-of-charge target values, said input training vectors and corresponding state-of-charge target values respectively established from values of a predetermined set of electro-chemical device parameters and corresponding values of state-of-charge resulting from electro-chemical device responses during predetermined charge cycling profile including charge cycling profiles corresponding to predefined vehicle driving schedules;

means for determing values of the predetermined set of electro-chemical device parameters during vehicle operation;

means for establishing an input recall vector from said values and an historically determined value of one of said electro-chemical device parameters;

processing means for propagating said input recall vector through said neural network to establish said state-of-charge of the electro-chemical device as a scalar value which non-asymptotically converges to values of 100% and 0% state-of-charge.

2. A neural network system as claimed in claim 1 wherein said neural network system comprises full inter-layer connections.

3. A neural network system as claimed in claim 1 wherein said neural network system comprises full inter-layer connection and no intra-layer connections.

4. A neural network system as claimed in claim 1 wherein said non-linear hidden layer processing elements comprise substantially sigmoidal transfer functions.

5. A neural network system as claimed in claim 1 wherein said predetermined set of electro-chemical device parameters comprise device current, cumulative device capacity discharged, a representative temperature of said device, and the minimum one of respective voltages of said plurality of electro-chemical modules.

6. A neural network system as claimed in claim 1 wherein said means for establishing said input vector comprises means for scaling said values.

7. A neural network system as claimed in claim 5 wherein said historically determined value of said one of said electro-chemical device parameters comprises the value of a previously determined minimum one of respective voltages of said plurality of electro-chemical modules.

8. A method for determining a state-of-charge of an electro-chemical device as claimed in claim 1 wherein said predetermined charge cycling profiles further includes constant current profiles.

9. A method for determining a state-of-charge of an electro-chemical device adapted for use in a motor vehicle for providing power to an electric machine in a drive mode and for receiving power from the electric machine in a regenerative mode, said electro-chemical device including a plurality of electro-chemical modules, the method comprising the steps:

providing a neural network architecture including a linear input layer, linear output layer and non-linear hidden layer fully connected with said input and output layers;

training the neural network by providing to the input layer a set of input training vectors including values of electro-chemical device parameters collected during charge cycling profiles corresponding to predefined vehicle driving schedules;

providing to the input layer a recall vector including values of substantially contemporaneously measured electro-chemical device parameters and an historically measured value of one of the electro-chemical device parameters; and

,

propagating the recall vector through the neural network to determine the state-of-charge of the electro-chemical device as a scalar value which non-asymptotically converges to values of 100% and 0% state-of-charge.

10. A method for determining a state-of-charge of an electro-chemical device as claimed in claim 9 wherein the electro-chemical device parameters comprise electro-chemical device current, cumulative electro-chemical device current discharged, voltage of the one of said plurality of electro-chemical modules having the minimum voltage from among said plurality of electro-chemical modules, temperature of said electro-chemical device, and the historically measured value of one of the electro-chemical device parameters comprises the value of a previously measured voltage of one of the plurality of electro-chemical modules having the minimum voltage.

11. A method for determining a state-of-charge of an electro-chemical device as claimed in claim 9 wherein said predetermined charge cycling profiles further include constant current profiles.

12. A method for determining a state-of-charge of an electro-chemical device as claim in claim 9 wherein said step of providing to the input layer a recall vector comprises the steps of filtering the values of the measured electro-chemical device parmeters.

13. A method for determining a state-of-charge of an electro-chemical device as claimed in claim 9 wherein said non-linear hidden layer comprises substantially signoidal transfer functions.

14. A method for determining a state-of-charge of an electro-chemical device as claimed in claim 13 wherein said substantially sigmoidal transfer functions are approximated by a fifth-order polynomial.

15. A method for determining a state-of-charge of an electro-chemical device adapted for use in a motor vehicle for providing power to an electric machine in a drive mode and for receiving power from the electric machine in a regenerative mode, said electro-chemical device including a plurality of electro-chemical modules, the method comprising the steps:

providing a neural network architecture including a plurality of N unity gain linear input processing elements, a unity gain linear output processing element for providing a scalar output representative of the state-of-charge, and a plurality of M substantially sigmoidal hidden layer processing elements, said hidden layer processing elements being fully connected with said input and output processing element;

training the numeral network by providing to the input layer a set of input training vectors including measurements of electro-chemical device parameters including (a) electro-chemical device current, (b) cumulative electro-chemical device current discharged, (c) voltage of the one of said plurality of electro-chemical modules having the minimum voltage from among said plurality, (d) temperature of said electro-chemical device, and (e) a most recent prior voltage of the one of said plurality of electro-chemical modules having the minimium voltage, all of which measurements having been collected during predetermined charge cycling profiles including constant current profiles and charge cycling profiles corresponding to predetermined vehicle driving schedules, and back-propagating respective states-of-charge corresponding to respective input training vectors; and

,

propagating a recall vector including filtered measurements of said elector-chemical device parameters through the neural network to determine the state-of-charge of the electro-chemical device as a scalar value which non-asymptotically converges to values of 100% and 0% state-of-charge.