基于MATLAB用LSTM源代码预测数据，程序能正常运行，结果却趋近于一条直线，不知道哪里出现的问题，下面是我写的代码，请给看一下。有偿
%%
clear
clc
%% 数据导入
load('G4_4ring.mat')
in = 10;%输入样本的数据量
out = 5;%输出样本的数据量
dataX = G4_4ring(1:200,:);
dataTrain =(dataX-mean(dataX))/std(dataX);
for i = 1 : 185
data_1(:,i) = dataTrain(i:i+9,1);
end
for i = 1 : 185
data_2(:,i) = dataTrain(i+10:i+14,1);
end
datatrain_in = data_1;%训练集输入
datatrain_out = data_2;%训练集输出
%% 设置BP参数
net.in = in;%输入节点
net.hid = 5;%隐藏节点
net.out = out;%输出节点
%权重偏置初始化
wf = 2*(rand(net.hid , net.hid+net.in) - 1/2);
bf = 2*(rand(net.hid,1) - 1/2);
wi = 2*(rand(net.hid , net.hid+net.in) - 1/2);
bi = 2*(rand(net.hid , 1) - 1/2);
wc = 2*(rand(net.hid ,net.hid+net.in) - 1/2);
bc = 2*(rand(net.hid , 1) - 1/2);
wo = 2*(rand(net.hid , net.hid+net.in) - 1/2);
bo = 2*(rand(net.hid , 1) - 1/2);
wy = 2*(rand(net.out , net.hid) - 1/2);
by = 2*(rand(net.out , 1) - 1/2);
%训练参数
learningrate = 0.001;
maxiter = 2;
iteration = 0;
%% 开始训练
for i = 1:maxiter
%% 前向计算
for t = 1:10
if t==1
Zx1 = wf*[zeros(out,1);datatrain_in(:,1)] + bf;%调整ht的维度
Zx2 = wi*[zeros(out,1);datatrain_in(:,1)]+bi;
Zx3 = wc*[zeros(out,1);datatrain_in(:,1)]+bc;
Zx4 = wo*[zeros(out,1);datatrain_in(:,1)]+bo;
for j=1:5
f_t(j,1) = Sigmoid(Zx1(j,:));%遗忘门
i_t(j,1) = Sigmoid(Zx2(j,:));
c_t_current(j,1) = tanh(Zx3(j,:));%输入门
o_t(j,1) = Sigmoid(Zx4(j,:));
end
c_t(:,1) = i_t(:,1).*c_t_current(:,1);%更新细胞状态
h_t(:,1) = o_t(:,1).*Tanh(c_t(:,1));%输出门
z_t = wy*h_t(:,1) + by;
y_t(:,1) = Sigmoid(z_t);
else
Zx1 = wf*[h_t(:,t-1);datatrain_in(:,t)]+bf;
Zx2 = wi*[h_t(:,t-1);datatrain_in(:,t)]+bi;
Zx3 = wc*[h_t(:,t-1);datatrain_in(:,t)]+bc;
Zx4 = wo*[h_t(:,t-1);datatrain_in(:,t)]+bo;
f_t(:,t) = Sigmoid(Zx1);%遗忘门
i_t(:,t) = Sigmoid(Zx2);
c_t_current(:,t) = Tanh(Zx3);%输入门
o_t(:,t) = Sigmoid(Zx4);
c_t(:,t) = f_t(:,t).*c_t(:,t-1)+i_t(:,t).*c_t_current(:,t);%更新细胞状态
h_t(:,t) = o_t(:,t).*Tanh(c_t(:,t));%输出门
z_t(:,t) = wy*h_t(:,t) + by;
y_t(:,t) = Sigmoid(z_t(:,t));
end
end
for t=1:10
Dby(:,t) = y_t(:,t) - datatrain_out(:,t);
sse_train = sumsqr(Dby);%误差平方和
end
fprintf('第 %d 次迭代 误差： %f\n' , i , sse_train);
iteration = iteration + 1;
%% 反向传播
for t = 10 : -1 : 2
if t==10
dby = y_t(:,t) - datatrain_out(:,t);
dwy = (y_t(:,t) - datatrain_out(:,t))*h_t(:,t)';
dh_t = wy'*dby;
dc_t = dh_t.*o_t(:,t).*(1-(Tanh(c_t(:,t))).^2);
dwo = dh_t.*Tanh(c_t(:,t)).*o_t(:,t).*(1-o_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwi= dc_t.*c_t_current(:,t).*i_t(:,t).*(1-i_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwf = dc_t.*c_t(:,t-1).*f_t(:,t).*(1-f_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwc = dc_t.*i_t(:,t).*(1-(c_t_current(:,t)).^2)*[h_t(:,t-1);datatrain_in(:,t)]';
dbo = dh_t.*Tanh(c_t(:,t)).*o_t(:,t).*(1-o_t(:,t));
dbi = dc_t.*c_t_current(:,t).*i_t(:,t).*(1-i_t(:,t));
dbf = dc_t.*c_t(:,t-1).*f_t(:,t).*(1-f_t(:,t));
dbc = dc_t.*i_t(:,t).*(1-(c_t_current(:,t)).^2);
%更新
Wf = dwf;
Wi = dwi;
Wc = dwc;
Wo = dwo;
Bf = dbf;
Bi = dbi;
Bc = dbc;
Bo = dbo;
Wy = dwy;
By = dby;
else
dby = y_t(:,t) - datatrain_out(:,t);
dwy =(y_t(:,t) - datatrain_out(:,t))*h_t(:,t)';
dh_t = wy'*dby+ wo(:,1:5)'*(dh_t.*Tanh(c_t(:,t+1)).*o_t(:,t+1).*(1-o_t(:,t+1))) + wi(:,1:5)'*(dc_t.*c_t_current(:,t+1).*i_t(:,t+1).*(1-i_t(:,t+1))) + wf(:,1:5)'*(dc_t.*c_t(:,t).*f_t(:,t+1).*(1-f_t(:,t+1))) + wc(:,1:5)'*(dc_t.*i_t(:,t+1).*(1-(c_t_current(:,t+1)).^2));
dc_t = dh_t.*o_t(:,t).*(1-(Tanh(c_t(:,t))).^2)+dc_t.*f_t(:,t+1);
dwo = dh_t.*Tanh(c_t(:,t)).*o_t(:,t).*(1-o_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwi = dc_t.*c_t_current(:,t).*i_t(:,t).*(1-i_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwf = dc_t.*c_t(:,t-1).*f_t(:,t).*(1-f_t(:,t))*[h_t(:,t-1);datatrain_in(:,t)]';
dwc = dc_t.*i_t(:,t).*(1-(c_t_current(:,t)).^2)*[h_t(:,t-1);datatrain_in(:,t)]';
dbo = dh_t.*Tanh(c_t(:,t)).*o_t(:,t).*(1-o_t(:,t));
dbi = dc_t.*c_t_current(:,t).*i_t(:,t).*(1-i_t(:,t));
dbf = dc_t.*c_t(:,t-1).*f_t(:,t).*(1-f_t(:,t));
dbc = dc_t.*i_t(:,t).*(1-(c_t_current(:,t)).^2);
%更新
Wf = Wf+dwf;
Wi = Wi+dwi;
Wc = Wc+dwc;
Wo = Wo+dwo;
Bf = Bf+dbf;
Bi = Bi+dbi;
Bc = Bc+dbc;
Bo = Bo+dbo;
Wy = Wy+dwy;
By = By+dby;
end
end
%% 更新
wf = wf+learningrateWf;
wi = wi+learningrateWi;
wc = wc+learningrateWc;
wo = wo+learningrateWo;
bf = bf+learningrate*Bf;
bi = bi+learningrate*Bi;
bc = bc+learningrate*Bc;
bo = bo+learningrate*Bo;
wy = wy+learningrate*Wy;
by = by+learningrate*By;
end
%% 测试
dataY = G4_4ring(201:250,:);
dataTest =(dataY-mean(dataY))/std(dataY);
for i = 1: 35
data_3(:,i) = dataTest(i:i+9,1);
end
for i = 1: 35
YTest(:,i) = dataTest(i+10:i+14,1);
end
%测试数据
datatest_in = data_3;%测试集输入
datatest_out = YTest;%测试集输出
%预测过程
for m = 1:35
if m==1
Zx1_T = wf*[zeros(out,1);datatest_in(:,m)]+bf;
Zx2_T = wi*[zeros(out,1);datatest_in(:,m)]+bi;
Zx3_T = wc*[zeros(out,1);datatest_in(:,m)]+bc;
Zx4_T = wo*[zeros(out,1);datatest_in(:,m)]+bo;
f_t_T(:,m) = Sigmoid(Zx1_T);%遗忘门
i_t_T(:,m) = Sigmoid(Zx2_T);
c_t_current_T(:,m) = tanh(Zx3_T);%输入门
o_t_T(:,m) = Sigmoid(Zx4_T);
c_t_T(:,m) = i_t_T(:,m).*c_t_current_T(:,m);%更新细胞状态
h_t_T(:,m) = o_t_T(:,m).*Tanh(c_t_T(:,m));%输出门
z_t_T(:,m) = wy*h_t_T(:,m) + by;
Y_t(:,m) =Sigmoid(z_t_T(:,m));
e_test(:,m) = Y_t(:,m) - datatest_out(:,m);%误差
else
Zx1_T = wf*[h_t_T(:,m-1);datatrain_in(:,m)]+bf;
Zx2_T = wi*[h_t_T(:,m-1);datatrain_in(:,m)]+bi;
Zx3_T = wc*[h_t_T(:,m-1);datatrain_in(:,m)]+bc;
Zx4_T = wo*[h_t_T(:,m-1);datatrain_in(:,m)]+bo;
f_t_T(:,m) = Sigmoid(Zx1_T);%遗忘门
i_t_T(:,m) = Sigmoid(Zx2_T);
c_t_current_T(:,m) = Tanh(Zx3_T);%输入门
o_t_T(:,m) = Sigmoid(Zx4_T);
c_t_T(:,m) = f_t_T(:,m).*c_t_T(:,m-1)+i_t_T(:,m).*c_t_current_T(:,m);%更新细胞状态
h_t_T(:,m) = o_t_T(:,m).*Tanh(c_t_T(:,m));%输出
z_t_T(:,m) = wy*h_t_T(:,m) + by;
Y_t(:,m) = Sigmoid(z_t_T(:,m));
e_test(:,m) = Y_t(:,m) - datatest_out(:,m);%误差
end
end
YPred = Y_tstd(dataY) + mean(dataY)
YTest = YTeststd(dataY) + mean(dataY);
XTest = datatest_in*std(dataY)+ mean(dataY);
%% 画图
figure('Position', [10, 10, 900, 400]);
S = XTest(1,:);
Y = [S;YPred];
D1 = XTest(1,:);
D2 = XTest(2:10,35)';
d = [D1,D2];
for i= 1 : 35
%figure('Position', [10, 10, 900, 400]);
plot(d(1:40),'k','LineWidth',2)
hold on
%plot(i+2O0:i+2O9,XTest(:,i),'y','LineWidth',2)
%plot(i+210:i+214,YTes(:,i), '.-k','LineWidth',2)
plot(i:i+5,Y(:,i), 'x-r')%
end
