基于LSTM神經網絡的電池SOH估算

??循環神經網絡（Recurrent Neural Network，RNN）與BP神經網絡不同，RNN網絡不僅考慮前一時刻的輸入，同時還賦予網絡對前面時刻信息的記憶能力。盡管RNN網絡具有較高的精度，但其存在著梯度消失的問題。對此，出現了一系列改進的RNN網絡，而LSTM神經網絡就是其中改進效果最好的一種。基于LSTM神經網絡的電池SOH估算方法具體如下所示：
??（1）鋰離子電池循環壽命數據
??數據來源于NASA研究中心搭建的鋰離子電池測試平臺，選取5號鋰離子電池（額定容量為2Ah）。循環測試實驗在室溫下進行：以1.5A的恒流電流為電池充電，直到充電截止電壓（4.2V），以恒壓電流充電，直到充電電流降至20mA；在2A恒流（CC）模式下放電，直至電池分別降至2.7V、2.5V、2.2V、2.5V。當電池達到EOL（end-of-life, EOL）標準時，實驗停止，額定容量下降30%。
??（2）數據預處理
??利用最大最小值歸一化方法將數據放縮到0和1之間。

??式中，xmax是輸入數據x的最大值，xmin是輸入數據x的最小值。
??（3）仿真分析
??將NASA鋰離子電池實驗數據中的充電容量作為模型輸入，模型的輸出為以放電容量為參考的電池 SOH。經過訓練，神經網絡可以迭代得到網絡權重及偏置具體參數。

??從圖中可以看出，模型對于電池容量衰減的整體趨勢預測是非常精準的。其具體代碼如下所示：

function pre_data = LSTM_main(d,h,train_data,test_data) %% 預處理 lag = 8; % d = 51; [train_input,train_output] = LSTM_data_process(d,train_data,lag); %數據處理[train_input,min_input,max_input,train_output,min_output,max_output] = premnmx(train_input',train_output');input_length = size(train_input,1); %樣本輸入長度 output_length = size(train_output,1); %樣本輸出長度 train_num = size(train_input,2); %訓練樣本個數 test_num = size(test_data,2); %測試樣本個數 %% 網絡參數初始化 % 結點數設置 input_num = input_length; cell_num = 10; output_num = output_length; % 網絡中門的偏置 bias_input_gate = rand(1,cell_num); bias_forget_gate = rand(1,cell_num); bias_output_gate = rand(1,cell_num); %網絡權重初始化 ab = 20; weight_input_x = rand(input_num,cell_num)/ab; weight_input_h = rand(output_num,cell_num)/ab; weight_inputgate_x = rand(input_num,cell_num)/ab; weight_inputgate_h = rand(cell_num,cell_num)/ab; weight_forgetgate_x = rand(input_num,cell_num)/ab; weight_forgetgate_h = rand(cell_num,cell_num)/ab; weight_outputgate_x = rand(input_num,cell_num)/ab; weight_outputgate_h = rand(cell_num,cell_num)/ab; %hidden_output權重 weight_preh_h = rand(cell_num,output_num); %網絡狀態初始化 cost_gate = 1e-6; h_state = rand(output_num,train_num+test_num); cell_state = rand(cell_num,train_num+test_num); %% 網絡訓練學習 for iter = 1:3000 %迭代次數yita = 0.01; %每次迭代權重調整比例for m = 1:train_num%前饋部分if(m==1)gate = tanh(train_input(:,m)' * weight_input_x);input_gate_input = train_input(:,m)' * weight_inputgate_x + bias_input_gate;output_gate_input = train_input(:,m)' * weight_outputgate_x + bias_output_gate;for n = 1:cell_numinput_gate(1,n) = 1 / (1 + exp(-input_gate_input(1,n)));%輸入門output_gate(1,n) = 1 / (1 + exp(-output_gate_input(1,n)));%輸出門%sigmoid函數endforget_gate = zeros(1,cell_num);forget_gate_input = zeros(1,cell_num);cell_state(:,m) = (input_gate .* gate)';elsegate = tanh(train_input(:,m)' * weight_input_x + h_state(:,m-1)' * weight_input_h);input_gate_input = train_input(:,m)' * weight_inputgate_x + cell_state(:,m-1)' * weight_inputgate_h + bias_input_gate;forget_gate_input = train_input(:,m)' * weight_forgetgate_x + cell_state(:,m-1)' * weight_forgetgate_h + bias_forget_gate;output_gate_input = train_input(:,m)' * weight_outputgate_x + cell_state(:,m-1)' * weight_outputgate_h + bias_output_gate;for n = 1:cell_numinput_gate(1,n) = 1/(1+exp(-input_gate_input(1,n)));forget_gate(1,n) = 1/(1+exp(-forget_gate_input(1,n)));output_gate(1,n) = 1/(1+exp(-output_gate_input(1,n)));endcell_state(:,m) = (input_gate .* gate + cell_state(:,m-1)' .* forget_gate)'; endpre_h_state = tanh(cell_state(:,m)') .* output_gate;h_state(:,m) = (pre_h_state * weight_preh_h)';%誤差計算Error = h_state(:,m) - train_output(:,m);Error_Cost(1,iter)=sum(Error.^2); %誤差的平方和（4個點的平方和）if(Error_Cost(1,iter)<cost_gate) %判斷是否滿足誤差最小條件flag = 1;break;else [ weight_input_x,...weight_input_h,...weight_inputgate_x,...weight_inputgate_h,...weight_forgetgate_x,...weight_forgetgate_h,...weight_outputgate_x,...weight_outputgate_h,...weight_preh_h ] = LSTM_updata_weight(m,yita,Error,...weight_input_x,...weight_input_h,...weight_inputgate_x,...weight_inputgate_h,...weight_forgetgate_x,...weight_forgetgate_h,...weight_outputgate_x,...weight_outputgate_h,...weight_preh_h,...cell_state,h_state,...input_gate,forget_gate,...output_gate,gate,...train_input,pre_h_state,...input_gate_input,...output_gate_input,...forget_gate_input,input_num,cell_num);endendif(Error_Cost(1,iter)<cost_gate)break;end end%% 測試階段 %數據加載 test_input = train_data(end-lag+1:end); test_input = tramnmx(test_input',min_input,max_input); % test_input = mapminmax('apply',test_input',ps_input);%前饋 for m = train_num + 1:train_num + test_numgate = tanh(test_input' * weight_input_x + h_state(:,m-1)' * weight_input_h);input_gate_input = test_input' * weight_inputgate_x + h_state(:,m-1)' * weight_inputgate_h + bias_input_gate;forget_gate_input = test_input' * weight_forgetgate_x + h_state(:,m-1)' * weight_forgetgate_h + bias_forget_gate;output_gate_input = test_input' * weight_outputgate_x + h_state(:,m-1)' * weight_outputgate_h + bias_output_gate;for n = 1:cell_numinput_gate(1,n) = 1/(1+exp(-input_gate_input(1,n)));forget_gate(1,n) = 1/(1+exp(-forget_gate_input(1,n)));output_gate(1,n) = 1/(1+exp(-output_gate_input(1,n))); endcell_state(:,m) = (input_gate .* gate + cell_state(:,m-1)' .* forget_gate)'; pre_h_state = tanh(cell_state(:,m)') .* output_gate;h_state(:,m) = (pre_h_state * weight_preh_h)';% 將當前預測點作為下一步輸入數據test_input = postmnmx(test_input,min_input,max_input);now_prepoint = postmnmx(h_state(:,m),min_output,max_output);%test_input = mapminmax('reverse',test_input,ps_input);test_input = [test_input(2:end); now_prepoint];test_input = tramnmx(test_input,min_input,max_input); endpre_data = postmnmx(h_state(:,train_num + h:h:train_num + test_num),min_output,max_output);all_pre = postmnmx(h_state(:,1:h:train_num + test_num),min_output,max_output);% 畫圖 figure title('LSTM預測') hold on plot(1:size([train_data test_data],2),[train_data test_data], 'o-', 'color','r', 'linewidth', 1); plot(size(train_data,2) + h:h:size([train_data test_data],2),pre_data, '*-','color','b','linewidth', 1); plot([size(train_data,2) size(train_data,2)],[-0.01 0.01],'g-','LineWidth',4); legend({ '真實值', '預測值'}); end

??想了解更過相關仿真，可以關注我的微信公眾號。

總結

以上是生活随笔為你收集整理的电池SOH仿真系列-基于LSTM神经网络的电池SOH估算方法的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。