How to design an AI-based Battery Management System (BMS) for EV State of Charge (SOC) estimation using MATLAB & Simulink?

State of Charge (SOC) estimation is a critical task in Electric Vehicle (EV) Battery Management Systems (BMS). Traditional methods like Coulomb counting suffer from current sensor drift, while open-circuit voltage (OCV) methods require long relaxation times. Using an AI model such as a Long Short-Term Memory (LSTM) network in MATLAB allows real-time, non-linear SOC tracking under dynamic driving conditions.

Step 1: Data Collection & Preparation

First, generate battery training data (voltage, current, temperature, and true SOC) using the Simscape Electrical Battery block or experimental discharge datasets (e.g., US06 or Federal Urban Driving Schedules).

% Load dynamic battery drive cycle dataset
data = load('EV_DriveCycle_Data.mat');
XTrain = [data.Voltage, data.Current, data.Temperature]'; % Inputs (3 x N)
YTrain = data.SOC';                                       % Target (1 x N)

% Normalize feature arrays
mu = mean(XTrain, 2);
sig = std(XTrain, 0, 2);
XTrainNorm = (XTrain - mu) ./ sig;

Step 2: Building the LSTM Network Architecture

Define an LSTM network in MATLAB Deep Learning Toolbox that takes continuous sequential time-series data and outputs estimated SOC percentages.

numFeatures = 3;
numHiddenUnits = 125;
numResponses = 1;

layers = [ ...
    sequenceInputLayer(numFeatures)
    lstmLayer(numHiddenUnits, 'OutputMode', 'sequence')
    reluLayer
    fullyConnectedLayer(numResponses)
    regressionLayer];

options = trainingOptions('adam', ...
    'MaxEpochs', 200, ...
    'GradientThreshold', 1, ...
    'InitialLearnRate', 0.005, ...
    'LearnRateSchedule', 'piecewise', ...
    'LearnRateDropPeriod', 50, ...
    'Verbose', 0);

% Train the neural network
bms_net = trainNetwork(XTrainNorm, YTrain, layers, options);

Step 3: Integrating with Simulink & Simscape

Export the trained network using the Stateful Predict block from Deep Learning Toolbox into your main Simulink BMS model. Route real-time sensor measurements (pack voltage, line current, module temperature) into the block to estimate SOC online.