put performance to zero
2 views (last 30 days)
Show older comments
Hello everybody, i created a feedforward network with the following code. I actually try to make prediction with this giving the network an input vector u1 with measured temperature and an interval avout [7,6°C - 22,8°C] and an output vector d1 with measured solar radiation and an interval about [0 W/m^2 - 1196 W/m^2]. These vectors have a lenght of around 10000 lines. If i train the network i start with an initial performance about 1,05e^5 and end around 4,64e^4 after the training. The error os calculated as MSE and lays around 40% from the initial value. Taking an input and output vector with just 30 lines I end with a rest-performace from 2% of the initial error. I actually thought the performace would be better the more data i use, but this is a contradiction to this. Someone got an idea why this training performance is that bad and how i can fix this to get better solutions? I included the plot from the given output (green) vector and the output of the network (red) so u'll see this a little bit better. I hope it helps a bit.
Thx anyway!!!
j=(1:1:10080);
r=xlsread('Mai_woche.xlsx');
u=r(1:10080,1);
u1=u';
y=xlsread('Mai_woche.xlsx');
d=y(1:10080,2);
d1=d';
k=10;
x=50;
y=1;
s=rng;
iweights=rand(k,y);
lweights=rand(x,k);
b1=rand(k,y);
b2=rand(x,y);
net=newff(minmax(u1),[k,x,y],{'tansig','tansig','purelin'},'trainlm');
net.IW{1,1}=iweights;
net.LW{2,1}=lweights;
net.b{1}=b1;
net.b{2}=b2;
net.IW{1,1}
net.LW{2,1}
net.trainParam.show = NaN;
net.trainParam.lr = 0.04;
net.trainParam.epochs = 10000;
net.trainParam.goal = 1;
net.trainParam.min_grad=1e-30;
net.trainParam.mu_max=1e18;
net.trainParam.mu_inc=20;
net.trainParam.mu_min=1e-5;
net.performParam.ratio=0.5;
net.trainParam.time=inf;
[net,tr]=train(net,u1,d1);
a=sim(net,u1);
e=d1-a;
perf=mse(e);
%3. Plot
plot(j,a,'r',j,d1,'g')
0 Comments
Answers (1)
See Also
Categories
Find more on Sequence and Numeric Feature Data Workflows in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!