TECHNICAL FIELD

[0001] The present invention relates to the field of battery management systems, and in particular, relates to a hybrid system and method for estimating the state of charge of a battery.

BACKGROUND

[0002] Background description includes information that may be useful in understanding the present invention. It is not an admission that any of the information provided herein is prior art or relevant to the presently claimed invention, or that any publication specifically or implicitly referenced is prior art.

[0003] The automotive industry is currently experiencing a paradigm shift from conventional, diesel, and gasoline-propelled vehicles into second-generation hybrid and electric vehicles. The rechargeable battery is the most critical component in an electric vehicle.

[0004] State of Charge (SOC) estimation is one of the most important functions of battery management systems (BMSs), which is defined as the percentage of the remaining charge inside the battery to its maximum capacity. SOC indicates when the battery needs to be recharged. Battery SOC estimation is required for many battery-management functions, for example, but not limited to, charge/discharge control, remaining useful time/ driving range predictions, and battery power capability estimations. As an essential performance indicator, the state of charge (SOC) reflects the residual capacity of a battery. To ensure the safe operation of systems, it is vital to estimate the battery SOC accurately. Inaccurate SOC estimations can lead to user dissatisfaction, mission failures, the inefficient performance of the vehicles, premature battery failures, etc. While a system or component exhibits degradation during its life cycle, there are various methods to predict its future performance and assess the time frame until it no longer performs its desired functionality.

[0005] Many methods are available to estimate the battery SOC, each with its own advantages and disadvantages. The current algorithm for the estimation of SOC uses a physics-based battery model for the prediction of SOC followed by a Kalman filter to correct the predicted SOC. State variables calculated by the battery model are the predictions and inputs to the Kalman filter. Corrected state variables obtained from the Kalman filter are the estimations. The Kalman filter is a special case of a ‘sequential probabilistic inference’, an estimation algorithm that corrects the predictions of state variables in the presence of process and sensor noise. The existing system may require retuning of the battery model/Kalman filter parameters for different load cycles. This makes it impractical for real-time deployment. Further, in physics-based methods, the major problem for real-time SOC estimation is the computational complexity of the coupled partial differential equations (PDEs) which are used to describe the physical processes inside the battery.

[0006] Electrochemical models have intrinsic advantages for SOC estimation since it can relate battery internal physical parameters, e.g. lithium concentrations, to SOC. However, the computational complexity of the electrochemical model is the major obstacle to its application in a real-time BMS. Currently, electrochemical mechanism-based models require huge computation effort for solving partial differential equations and equivalent circuit models do not consider the actual battery mechanism.

[0007] The current machine learning models are lacking generalization ability due to their data-driven nature. Further, the machine learning models are sensitive to the amount and quality of training data. All these limitations hinder the model's effectiveness in estimating the SOC of the battery.

[0008] Thus, there is a need for an efficient, robust, and improved battery SOC estimation system and method, which utilizes and combines the advantages offered by the physics-based model and the machine learning model for accurately estimating SOC of the battery without requiring recalibration or retuning of the battery model or the Kalman filter parameters, while overcoming the limitations of each of these models.

OBJECTS OF THE INVENTION

[0009] An object of the present invention is to provide an efficient, improved, and robust system and method for accurately estimating the state of charge of the battery over a wide range of dynamic data.

[0010] Another object of the present invention is to provide a hybrid system and method for SOC estimation that combines a physics-based battery model and a data-driven model to overcome the limitations of each of these individual models

[0011] Yet another object of the present invention is to provide a hybrid system and method for SOC estimation that eliminates the need for Kalman filter by replacing it with a data-driven neural network, thereby, eliminating the need for recalibration of the Kalman filter. [0012] Yet another object of the present invention is to provide a hybrid system and method for battery SOC estimation that does not require to retune the battery model.

[0013] Still another object of the present invention is to provide a hybrid system and method for battery SOC estimation which improves the accuracy of SOC estimation.

SUMMARY

[0014] The present invention relates to an efficient, robust, and improved system and method, which utilizes and combines the advantages offered by the physics-based model and the machine learning model for accurately estimating the SOC of the battery without requiring recalibration or retuning of the battery model or the Kalman filter parameters. The battery can be associated with a hybrid vehicle, electric vehicle, inverter, gadgets, consumer devices, industrial machines, and appliances, and the likes.

[0015] According to an aspect, the hybrid system and method of the present disclosure comprises a physics-based battery model, and a neural network (NN) computing unit comprising a plurality of neural networks for accurately estimating the SOC of the battery. The battery model generates output based on the values of current, voltage, temperature, and state of health indicators of the battery, which is then provided to the NN computing unit, which eliminates the need of conventional Kalman filters. The neural networks of the hybrid system and method of the present invention predict and correct two state variables: a state of charge of the battery cell (SOC) and a polarization voltage (Vp) associated with the battery. Through closed-loop feedback, the corrected polarization voltage (Vp) and the state of charge of the battery cell (SOC)being predicted by the NN computing unit are fed back to the battery model, which greatly improves the accuracy of the SOC estimation by the proposed system and method.

[0016] In an aspect, one of the neural networks associated with the battery determines the corrected Vp by processing the values of voltage prediction error and Vp being predicted by the battery model, and the values of the temperature and normalized current of the battery. Further, another neural network associated with the NN computing unit determines the corrected SOC by processing values of the SOC, Vp, and normalized voltage of the battery being predicted by the battery model, and normalized value of the Vcell of the battery.

[0017] In another aspect, a single neural network may be trained to determine the corrected SOC and the Vp based on the SOC, Vp, and other electrical attributes of the battery being predicted by the battery model. [0018] In an aspect, the NN computing unit is initially trained to determine optimal values of the SOC and the Vp using a training dataset comprising state variables, and intermediate parameters received from a well-tuned filter model, before configuration of the proposed system with a battery for estimation of the SOC of the battery.

[0019] Accordingly, the proposed hybrid system and method utilize and combine the advantages offered by the physics-based model and the machine learning model for accurately estimating SOC of the battery without requiring recalibration or retuning of the battery model or any Kalman filter, and offering advantages of improved accuracy and increased robustness to retain this accuracy over a wide range dynamic data.

BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification. The drawings illustrate exemplary embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0021] FIG. 1A illustrates an exemplary block diagram of the proposed hybrid system for estimating the SOC of a battery according to an embodiment of the present invention.

[0022] FIG. IB illustrates an exemplary architecture of the proposed system of FIG. 1A.

[0023] FIG. 2 illustrates an exemplary equivalent circuit diagram of the battery model of the proposed system and method according to an embodiment of the present invention.

[0024] FIGs. 3A and 3B illustrate an exemplary structure of the first and second neural networks of the proposed system and method for estimating the state of charge (SOC) and polarization voltage (Vp) of the battery according to an embodiment of the present invention.

[0025] FIG. 4 illustrates steps involved in the proposed hybrid method for estimating the SOC of a battery according to an embodiment of the present invention.

[0026] FIG. 5 illustrates an exemplary workflow of the proposed hybrid method for estimating the SOC of a battery according to an embodiment of the present invention.

[0027] FIGs. 6 to 27 illustrate various experimental graphs for validation of the proposed system and method.

DETAILED DESCRIPTION

[0028] The present invention relates to a system and method for battery state-of-charge (SOC) estimation and in particular, relates to a hybrid system for estimating SOC of a battery. According to an aspect, the proposed system comprises a physics-based battery model equivalent to the battery, wherein the battery model is configured to: receive a first set of data packets comprising values of current, voltage, temperature, and state of health indicators of the battery; and predict values of state of charge (SOC), and one or more electrical attributes associated with the battery based on the received first set of data packets, and correspondingly generate a second set of data packets. Further, the system comprises a neural network computing unit operatively coupled to the battery model, wherein the neural network computing unt comprises one or more processor operatively coupled to a memory storing one or more instructions being executable by the one or more processors. The NN computing unit is configured to receive the second set of data packets, and correspondingly extract the predicted SOC, and the predicted one or more electrical attributes of the battery, and process the predicted SOC and the predicted one or more electrical attributes to determine a corrected SOC, and a corrected polarization voltage (Vp) indicative of the SOC, and the Vp of the battery. The neural network computing unit is operatively coupled to the battery model in a closed loop feedback to tune the battery model based on the corrected SOC and the corrected Vp.

[0029] In an embodiment, the one or more electrical attributes comprises series resistance (Rs), voltage prediction error, parallel resistance (Rp), parallel capacitance (Cp), open circuit voltage (OCV), and terminal voltage (Vcell) of the battery. The state of health indicators of the battery comprises cell capacity factor (SOHC), and cell resistance factor (SOHR) of the battery

[0030] In an embodiment, a first neural network of the neural network computing unit is configured to process the values of the voltage prediction error and the Vp being predicted by the battery model, and the values of the temperature and normalized current of the battery, to determine the corrected Vp. Further, a second neural network of the neural network computing unit is configured to process the values of the SOC, the Vp, and normalized voltage of the battery being predicted by the battery model, and normalized value of the Vcell of the battery, to determine the corrected SOC.

[0031] In an embodiment, the neural network computing unit is trained to determine optimal values of the SOC and the Vp using a training dataset comprising state variables, and intermediate parameters received from a well-tuned Kalman filter model. The neural network computing unit comprises at least one feed forward neural network comprising one or more hidden layers, each having a predefined number of nodes, wherein a positive linear transfer function is used as an activation function for the nodes in the one or more hidden layers, and a pure linear function is used at an output node of the one or more hidden layers.

[0032] In an embodiment, the neural network computing unit is configured to dynamically adapt the battery model based on the first set of data packets, and previously determined values the corrected SOC and the corrected Vp to determine subsequent updated values of the one or more electrical attributes, and subsequent corrected values of the SOC, and the Vp.

[0033] According to another aspect, the present disclosure elaborates upon a method for estimating state of charge (SOC) of a battery. The method comprising a step of receiving, by a physics-based battery model equivalent to the battery, a first set of data packets comprising values of current, voltage, temperature, and state of health indicators (SOH) of the battery. The method further comprises a step of predicting, by the battery model, values of state of charge (SOC), and one or more electrical attributes associated with the battery based on the received first set of data packets, and correspondingly generating a second set of data packets. Further, the method comprises the steps of receiving, by a neural network computing unit operatively coupled to the battery model, and comprising one or more processors, the second set of data packets, and correspondingly extracting the predicted SOC, and the predicted one or more electrical attributes of the battery. Furthermore, the method comprises a step of processing, by the neural network computing unit, the predicted SOC and the predicted one or more electrical attributes to determine a corrected SOC, and a corrected polarization voltage (Vp) indicative of the SOC, and the Vp of the battery.

[0034] Referring to FIG. 1A and IB, in an aspect, the proposed system 100 for estimation of SOC of a battery comprises a physics-based battery model 102 equivalent to the battery, which is operatively coupled to a neural network (NN) computing unit 104 comprising a plurality of neural networks 104-1 and 104-2 for accurately estimating SOC of the battery. The battery model 102 is configured to receive a first set of data packets comprising current, voltage, temperature, and state of health indicators of the battery. Further, battery model 102 predicts values of state of charge (SOC), and one or more electrical attributes associated with the battery as outputs, based on the received first set of data packets, and correspondingly generates a second set of data packets.

[0035] The generated output (second set of data packets) of the battery model 102 is then provided to the NN computing unitl04 for further prediction and correction of two state variables: a state of charge of the battery cell (SOC) and a polarization voltage (Vp) associated with the battery, which eliminates the need of conventional Kalman filters. The neural networks 104-1 and 104-2 of the NN computing unit 104 extract and process the SOC, and electrical attributes predicted by the battery model 102 to determine the corrected SOC and the Vp. The neural networks 104-1 and 104-2 are connected to the battery model 102 in a closed-loop feedback to tune the battery model 102 so that the corrected Vp and the SOC being predicted by the NN computing unit 104 are fed back to battery model 102, which greatly improves the accuracy of the SOC estimation by the proposed system 100.

[0036] In an embodiment, the proposed system 100 is used for accurate estimation of the state of charge of the Li-ion batteries used in hybrid/electric vehicles, but not limited to the likes. Additionally, the proposed system lOOis used for estimating the state of charge of various types of batteries used in various applications, including, but not limited to, hybrid vehicle, electric vehicle, inverter, industrial applications, consumer gadgets, and consumer appliances.

[0037] In an exemplary embodiment, as illustrated in FIG. 2, the electrical attributes of the battery or battery model 100 comprise series resistance (Rs), terminal voltage (Vcell) of the battery voltage prediction error (measured Vcell-predicted Vcell), parallel resistance (Rp), parallel capacitance (Cp), and open-circuit voltage (OCV). The state of health indicators of the battery comprises a cell capacity factor (SOHC), and cell resistance factor (SOHR) of the battery.

[0038] The behavior of the battery (also referred to as cell herein), which is an electrochemical power source, can be understood by knowing the internal chemical process. This behavior is represented electrically as the physics-based battery model in the simplest form using a voltage source. But the actual cell voltage depends on several variables such as load current, State of charge, internal impedance which in turn are dependent on temperature and past usage. When loaded, the cell terminal voltage Vcell drops slowly. This deviation of cell terminal voltage Vcell from the open-circuit voltage (OCV) is known as polarization voltage Vp. This is modeled in the battery model by using an internal resistance Rs. When the battery is disconnected from the load and allowed to rest, its terminal voltage does not return to open-circuit voltage immediately, rather decays gradually. The slowly occurring diffusion process in the cell is the reason behind this diffusion voltage phenomenon. This phenomenon is modeled using one or more parallel Resistor-Capacitor (Rp, Rp) sub-circuits.

[0039] As seen, the battery model 102 comprises of a dependent voltage source (OCV) modeled as a function of SOC, Series resistance element Rs, parallel R-C network with R _P as parallel resistance, and Cpas parallel capacitance. Rs, Rp, and Cp have been represented as functions of SOC and temperature, which is being derived from experimental data obtained by performing a Hybrid Pulse Power Characterization (HPPC) test on a cell, using one or more sensors 106.

[0040] The cell capacity factor (SOHC) accounts for the degradation of cell capacity from its normal capacity, and cell resistance factor (SOHR) accounts for an increase in cell resistance over period usage, which are the real-time indicator of SOH of the battery. Any sensors error and initial SOC estimation by the battery model 102, may bring inaccuracy in SOC estimation, however, the NN computing network 104 corrects the SOC estimated by the battery model 102.

[0041] In an embodiment, as illustrated in FIG. IB, the NN computing unit 104comprises one or more processors 110 operatively coupled to a memory 112 storing one or more instructions being executable by the one or more processors 110 to receive the outputs from the battery model 102 and process them to estimate or predict the corrected SOC, and Vp. In an embodiment, the neural networks 104-1 and 104-2 of the NN computing unit 104 are feed-forward neural networks comprising one or more hidden layers, each having a predefined number of nodes. A positive linear transfer function is used as an activation function for the nodes in the one or more hidden layers, and a pure linear function is used at an output node of the one or more hidden layers for each neural network 104-1 and 104-2.

[0042] Referring to FIG. 3A, the NN computing unit 104 comprises a first neural network 104-1 configured to receive and process the values of the voltage prediction error and the Vp being predicted by the battery model 102, and the values of the temperature and normalized current (is an input feature derived using the combination of current, internal resistance (Rs) of the battery and the battery capacity in order to compensate parametric differences in batteries of same chemistry) of the battery, to determine the corrected Vp. The first neural network 104-1 is a feed-forward neural network that has two hidden layers with 5 nodes in each layer. A positive linear transfer function is used as an activation function for nodes in both the hidden layers and the pure linear transfer function at the output node.

[0043] Referring to FIG. 3B, a second neural network 104-2 of the neural network computing unit 104 is configured to receive and process the values of the SOC, the Vp, and normalized voltage of the battery being predicted by the battery model 102, and the normalized value of the Vcell of the battery, to determine the corrected SOC. The second neural network 104-2 is a feed-forward neural network having 2 hidden layers with 4 nodes in each layer. A positive linear transfer function is used as an activation function for nodes in both the hidden layers and pure linear transfer function at the output node. [0044] In an embodiment, the prediction unit 114 of the NN computing unit 104 enables the processors 110 to receive the generated output (second set of data packets) of the battery from the battery model 102, and further enable each of the neural networks 104-1 and 104-2 to extract and process the SOC, and electrical attributes predicted by the battery model 102 to determine the corrected SOC and the Vp.

[0045] In an embodiment, system 100 comprises sensors 106 operatively coupled with the battery to monitor current, voltage, temperature, and state of health indicators of the battery, and correspondingly generate and transmit the first set of data packets to the battery model 102. In an exemplary embodiment, the sensors comprise a current sensor, voltage sensor, temperature sensor, and SOH monitoring sensors or systems, but not limited to the likes.

[0046] In an embodiment, the training unit 116 of the system 100 enables the NN computing unit 104 to receive a training dataset comprising state variables, and intermediate parameters received from a well-tuned Kalman filter model, and further enable the processors 110 to train the first neural network 104-1, and the second neural networkl04-2 to determine optimal values of the Vp, and the SOC, respectively. In another embodiment, the training unit 116 may enable the processors 110 to train a single neural network using the training dataset comprising state variables, and intermediate parameters received from a well-tuned Kalman filter model to determine optimal values of the Vp, and the SOC.

[0047] The training dataset, as well as the, received set of first data packets, corresponding values of the electrical attributes, predicted values of SOC, Vp, and corrected values of the SOC and the Vp, are stored in a database 118, which further allows the system 100 to tune the battery model 102 if required, and also train the NN computing unit 104 for improving the accuracy of the system 100.

[0048] It is to be appreciated by a person skilled in the art that the suggested Kalman filter model is neither a component of the proposed system nor used during the operation of the proposed system. The Kalman filter is only used initially, prior to implementation/configuration of the proposed system with a battery for estimation of battery SOC, just for training the NN computing unit 104 of the proposed system 100. The elimination of Kalman filter model for SOC estimation in the proposed system helps overcome the drawbacks, complexities, and limitations associated with the Kalman filter, and makes the present invention efficient, accurate, and robust.

[0049] In an embodiment, the proposed system 100 is configured to estimate the state of health (SOH) of the battery. Degradation in the health of a battery is modeled in a two-fold manner as degradation in cell capacity and degradation in cell power. The capacity fade is fading of energy storage capacity of the battery due to aging and usage, which is expressed as SOHC = ^CapActual _> where, Cap _Actuat is actual aged cell capacity and Cap _nom is rated cell aPnom capacity J of a fresh cell ,,

[0050] The power fade is fading of the electrical power supplying capacity of the battery, due to an increase in the internal resistance of the battery due to aging and usage. Power fade is expressed as a ratio of actual cell internal resistance to fresh cell internal resistance, which is expressed as, SOHR = ^Rs - ^Actual s_fresh_cell where, R _s Actual is actual cell internal resistance of an aged cell and R _s j _r esh_ceii is the internal resistance of the fresh cell, and/? _s Actual =

[0051] Referring to FIG. 4, in another aspect, the present invention elaborates upon a method 400 for estimating the SOC of a battery. Method 400 involves a physics-based battery model 102 equivalent to the battery, which is operatively coupled to a neural network (NN) computing unit 104 comprising a plurality of neural networks 104-1 and 104-2for accurately estimating the SOC of the battery. Method 400 comprises step 402 of receiving, by the battery model 102, a first set of data packets comprising values of current, voltage, temperature, and state of health indicators (SOH) of the battery being monitored or sensed by sensors 106 connected to the battery. Method 400 comprises step 404 of predicting, by the battery model 102, values of state of charge (SOC), and one or more electrical attributes associated with the battery based on the received first set of data packets generated by the battery model in step 402, and correspondingly generating a second set of data packets.

[0052] The method 400 further comprises step 406 of receiving, by the neural network computing unit 104 that is connected to the battery model 102 in a closed-loop feedback, the second set of data packets generated in step 404, and correspondingly extracting the predicted SOC, and the predicted one or more electrical attributes of the battery. Further, method 400 comprises step 408 of processing, by the neural network computing unit 104, the predicted SOC and the predicted one or more electrical attributes to determine a corrected SOC, and Vp indicative of actual SOC, and the Vp of the battery.

[0053] In an embodiment, method 400 further comprises step 410 of dynamically adapting the battery model 102 based on the first set of data packets and previously determined values the corrected SOC and the corrected Vp to determine subsequently updated values of the one or more electrical attributes, and subsequently corrected values of the SOC, and the Vp.

[0054] Referring to FIG. 5, an exemplary workflow 500 of the proposed method is disclosed. The workflow involves step 502 of obtaining the Hybrid Pulse Power Characterization (HPPC) test data of the battery, using sensors. Based on the HPCC test data the battery model 102 is tuned at step 504, and data profiles/subsets for training the NN computing unit is done at step 506. Further, a Kalman filter (which is not used during the actual implementation of the proposed method) is tuned at step 508 based on the data profiles/subsets of step 504. Finally, the state variables and intermediate parameters generated by the well-tuned Kalman filter are taken by the processors 110 of the present invention 100 and 400 at step 510 to train the neural networks 104-1 and 104-2 of the NN computing unit 104 at step 512 to determine optimal values of corrected SOC and Vp.

[0055] Once, the NN computing network 104 is trained, the Kalman filter is replaced by the NN computing unit at step 514, and correspondingly the NN computing unit 104 is validated using untrained and unseen data at step 516. At step 516, the performance of the NN computing unit 104 is validated. If the performance of the NN computing network 104 is not found satisfactory, the size of the training dataset is increased, and the NN computing unit 104 is again trained using the training dataset taken from the well-tuned Kalman filter as in step 512. Further, if the performance of the NN computing network is found satisfactory, the NN computing unit is deployed at step 520 in the proposed system with the battery to validate the performance, eliminating the need for the Kalman filter or recalibration during the actual operation of the proposed system and method.

[0056] In an implementation, the proposed system and method were validated by implementing them on Lithium Cobalt Oxide (LCO) chemistry data, lithium ferrophosphate (LFP) chemistry, and Lithium Nickel Manganese Cobalt Oxide (NMC) chemistry data. The 3 data sources chosen to validate the implementation differ from each other with respect to cell chemistries. Within each dataset, there are subsets of data with varying charge-discharge patterns or types of drive cycles.

[0057] Cell terminal voltage (Vcell) is a measurable physical quantity that is used as reference cell voltage. The terminal cell voltage is an output variable that is dependent on state variables V _P and SOC. Here, the estimated Vcell has been calculated using variables corrected by the NN computing unit. Value of OCV is obtained from OCV-SOC characteristics corresponding to corrected SOC by NN computing unit 104, and Vp is NN computing unit corrected polarization voltage. Thus, accurate estimation of Vp & SOC leads to accurate estimation of cell voltage. Therefore, an error between reference cell voltage and estimated cell voltage has been used to validate the performance of the proposed design.

VALIDATION 1

[0058] The LCO battery usage data considers battery cells that are subjected to varying charge/discharge cycles with randomly generated current profiles. These profiles, called

Random Walks (RW), provide current, voltage, and temperature values of a battery cell which undergoes various profiles of a series of charge and discharge cycles. Each profile has been run on a set of 4 batteries. In order to validate model performance on different charge/discharge patterns 3 profiles are selected which include a total of 11 RWs, and validation tests were conducted in MATLAB environment.

Table: 1 Validation 1 Dataset details

[0059] The battery model 102 has been tuned to represent the behavior of the cell type considered in the dataset. Cell OCV-SOC characteristic has been set to meet this cell specification as shown in FIG. 6. Kalman filter (KF) has been tuned independently to give satisfactory results for Profiles RWs resulting in 2 calibrations for the entire dataset as mentioned in below Table 2.

Table: 2 Validation 1 Training Dataset details [0060] Referring to FIG. 7, a current and temperature profile of LCO dataset profile 1 is shown, where a randomized pulse sequence of the discharge current is followed by charging for a randomly selected duration. Further, referring to FIG. 8, the current and temperature profile of second RW profile 2 of LCO data is shown, where randomized pulse sequence of the discharge current is followed by charging till maximum terminal cell voltage is reached (4.2V). Furthermore, referring to FIG. 9, current and temperature profile of LCO dataset RW Profile 3 is shown, where the cell is continuously operated using a randomized sequence of charging and discharging currents between -4.5A and 4.5A. Each voltage, Current, and temperature field of input data is pre-filtered using a moving average filter. A rationality check filter has also been applied to prohibit SOC estimation when input data is missing for a duration greater than the specified threshold.

VALIDATION 2

[0061] Validation 2 is performed on production data obtained from numerous drive cycle tests. Details about the dataset are given in table 5 below.

Table 3: Validation 2 dataset details

[0062] Referring to FIG. 10, a normalized OCV-SOC relationship tuned for these NMC cells has been shown. The dataset includes WLTC drive cycles tested at 14degC and 25degC. The dataset also includes non-standard drive cycles tested at different temperatures ranging from -30 degC to 25 degC. The battery model 102 and KF have been tuned for optimum performance of selected drive cycles leading to multiple calibrations. The NN computing unitl04 is trained using subsets of data from these drive cycles with adequate randomness such that NN computing unitl04 can predict outcomes effectively over entire data. Details of the training dataset are given in table 4 below.

Table 4: Validation2 training dataset details [0063] Validation 2 is done on actual vehicle production data to verify the efficacy of the approach on random, dynamic data with cell degradation impact and to check algorithm performance on real -world data.

RESULTS: Validation 1

[0064] The proposed system has been tested on the aforementioned 3 datasets. All the 3 datasets also provide cell actual capacity reference data. The Cell capacity factor input used in this implementation has been verified to be equal to the reference capacity value with a tolerance of ±5% deviation. Thus, the impact of capacity degradation has been effectively considered during the testing and validation.

The results obtained for the LCO dataset from KF based implementation and hybrid implementation are compared for the number of voltage estimation error occurrences above5% threshold of 0.05V. These details are tabulated below in Table 5.

Table 5: Comparison of error occurrences of voltage estimation error occurrence above 0.05V

[0065] A comparison between reference cell voltage and estimated cell voltage is shown in FIGs. 11 to 13. FIG. 11 illustrates the Reference and Estimated cell voltage v/s Time graph for LCO dataset Profilel cell3 (RW7). FIG. 12 illustrates the Reference and Estimated cell voltage v/s Time graph for LCO dataset Profile3 celll (RW9). FIG. 13 illustrates the Reference and Estimated cell voltage v/s Time graph for LCO dataset Profile2 celll (RW3). It can be seen from the results of FIGs. 11 to 13 that the voltage error between reference cell voltage and estimated cell voltage is within acceptable limits. Hence, it becomes evident that the proposed implementation estimates both the state variables V _P and SOC accurately to produce V _ce u output with the required tolerance.

[0066] Table 6 given below shows the average error in capacity estimated by the implementation with respect to reference cell capacity for all the LCO RW datacycles. Table 6: Capacity estimation error

[0067] FIG. 14 shows Estimated cell capacity and Reference cell capacity v/s time for LCO dataset Profile2 cell2 (RW4). FIG. 15 shows cell capacity estimation error v/s time for the LCO dataset Profile2 cell2 (RW4). Referring to FIGs. 14 and 15, a comparison of cell capacity estimated by the implemented model and reference cell capacity for LCO dataset profile2 cell2(RW4), and errors between the mare illustrated.

RESULTS: Validation 2

[0068] Validation results of testing on normalized NMC production data are tabulated in Table 7. Validation results are presented here as relative values and not the actual value. Table 7: MAE and standard deviation in Voltage estimation for NMC data

[0069] The graphs of the voltage estimation error of selected drive cycles used entirely for validation are shown in FIGs. 16 to 23, along with respective current profiles for representation purposes. FIG. 16 illustrates the current profile for WLTC drive cycle 1 at 25degC. FIG. 17 illustrates Voltage estimation error v/s Time for WLTC drive cycle 1 at 25degC.FIG. 18 illustrates the current profile for WLTC drive cycle 2 at 14degC.FIG. 19 illustrates Voltage estimation error v/s Time for WLTC drive cycle 2 at 14degC.FIG. 20 illustrates the current profile for Drive cycle 3 with startup SOHC= 80%. FIG. 21 illustrates Voltage estimation error v/s Time for Drive cycle 3 with startup SOHC= 80%.FIG. 22 illustrates the current profile for Cranking cycle 5 at 14degC. FIG. 23 illustrates Voltage estimation error v/s Time for Cranking cycle 5 at 14degC.

[0070] From the obtained results, it is seen that the trained NN computing unit 104 can perform SOC estimation accurately for all the data cycles and can effectively replace KF.SOHC and SOHR estimation errors with respect to reference SOHC and SOHR are tabulated in table 11 below. Further, referring to FIGs. 24 and 25, graphs for SOHC and SOHR comparison with reference data are shown.

Table 8:

HARDWARE VALIDATION DETAILS

[0071] After validating the NN computing unit-based design approach in the MATLAB environment then this was validated on MPC5746R Power Architecture MCU from NXP Semiconductors.

Table 9: Hardware validation setup details.

[0072] There were 2 SW design approaches considered while validating this NN on Controller. In approach I, the Single NN block evaluates all 15 Cell SOC and Pack SOC in a single execution with a periodic sample rate of 100msec. However, it was observed that with this approach NN could not process all the 15 cell SOC and pack SOC. So, an alternative approach II was designed where discrete NN blocks are used for single cell SOC evaluation with a periodic sample rate of 10msec. Thus, all the cells are evaluated in a period of 15 execution cycles. It was observed that NN could process all cell SOC and pack SOC with this approach. So, all the cell SOC and Pack SOC will be available for CAN transmission in 150msec which is less than the standard maximum permissible signal transmission rate of 200msec. Table 10: Comparison of RAM and ROM consumption by KF and NN SW design

Approaches in bytes

From the above table, it can be inferred as the NN approach with discrete NN blocks for each cell fairs better than the KF approach in ROM, RAM and Stack consumption.

Table 11: Hardware validation results summary.

[0073] It is to be appreciated by a person skilled in the art that by using the right set of input parameters, the architecture of the neural networks 104-1 and 104-2 of the NN computing unit 104 of proposed system 100 has been made analogous with Kalman filter (KF) framework. This NN computing unit 104 replaces and replicates the functionality of conventional KF, thereby eliminating the efforts of re-tuning the KF. Thus, the NN computing unit 104 trained using datasets obtained from multiple drive cycles and battery usage patterns can be employed for real-time battery management (BMS) applications.

[0074] FIG. 26 shows the number of error instances in SOC estimation by KF based and NN- based hybrid approach of the present invention. As seen, the number of voltage error instances over the permitted threshold is considerably reduced with the proposed hybrid system. Further, FIG. 27shows the % improvement in MAE of voltage estimation validated on LFP dataset using hybrid approach over conventional KF based approach. Thus, it is concluded thatthe proposed approach improves the SOC estimation accuracy. Furthermore, FIG. 27shows % improvement in MAE of voltage estimation of proposed hybrid approach over KF based approach.

[0075] Accordingly, the hybrid system and method of the present invention greatly improves the accuracy of the SOC estimation. Since the hybrid system and method of the present invention comprise both, physics-based battery model and a data-driven neural network model, it overcomes the disadvantages and limitations of each of these individual models. The hybrid system and method of the present inventioneliminate the need for a conventional Kalman filter by replacing it with a plurality of neural networks; thereby, eliminating the need for recalibration of the Kalman filter. Further, the hybrid system and method of the present invention do not require retuning the battery model.

ADVANTAGES OF THE PRESENT INVENTION

[0076] The present invention provides an efficient, improved, and robust system and method for accurately estimating the state of charge of the battery over a wide range of dynamic data.

[0077] The present invention provides a hybrid system and method for SOC estimation that combines a physics-based battery model and a data-driven model to overcome the limitations of each of these individual models.

[0078] The present invention provides a hybrid system and method for SOC estimation that eliminates the need for a Kalman filter by replacing it with a data-driven neural network, thereby, eliminating the need for recalibration of the Kalman filter.

[0079] The present invention provides a hybrid system and method for battery SOC estimation that does not require retuning the battery model.

[0080] The present invention provides a hybrid system and method for battery SOC estimation which improves the accuracy of SOC estimation.