Neural Network to predict weather not converging

Hi,

I developed a code to predict weather using neural network without using toolbox. I used the following paper to develop the algorithm:

An artificial neural network hourly temperature forecaster with applications in load forecasting

The problem I am facing in this code is it is not converging:

%-------------------------------------------%
% Reading temperature data
%-------------------------------------------%
filename='STL_2010.csv';
Temp_actual=csvread(filename,1,3,[1, 3, 8759, 3]);
Temp=Temp_actual.*0.001;

%-------------------------------------------%
% Neural Network
%-------------------------------------------%

%Initial temperatures
i=1;
temp24=Temp(i:i+23);
 [fut_min_t,min_index]= min(temp24);
 [fut_max_t,max_index]= max(temp24);
tempn24=Temp(i:i+23);

min_t_hat_p=fut_min_t;   
max_t_hat_p=fut_max_t;

%Neural Network weight initialization
v=rand(28,60);
w1=rand(60,60);
w2=rand(60,60);
w3=rand(60,24);
B=ones(60,24);
k_v=0.0001;
alpha1=0.001;
alpha2=0.001;
alpha3=0.001;
% temp_hat(i:i+23)=rand(1,24);
%Adaptive Scaling
syms m b;

flag=0;
for i=24:24:8736
   temp24=Temp(i:i+23);

     past_min_t=fut_min_t;
     past_max_t=fut_max_t;

     fut_min_t= min(temp24);
     fut_max_t= max(temp24);

     z=[past_min_t;past_max_t;fut_min_t;fut_max_t;tempn24];   %Neural Network Input

     %Neural Network
     phi1=tanh(v'*z);
     phi2=tanh(w1'*phi1);
     phi3=tanh(w2'*phi2);
     temp_hat(i:i+23)= w3'*phi3;          %tanh(w2'*phi2);

      error(i:i+23)=temp_hat(i:i+23)-Temp(i:i+23)';
   %Adaptive scaling
    %Formula: T_scaled=m*T_neural_o/p+bais

     min_index_p=min_index;
     max_index_p=max_index;

     min_t_hat=min(temp_hat(i:i+23));
     max_t_hat=max(temp_hat(i:i+23));
     [min_t,min_index]=min(Temp(i:i+23));
     [max_t,max_index]=max(Temp(i:i+23));
     min_index=min_index+i;
     max_index=max_index+i;
     if min_index_p < max_index_p

         [solm, solb] = solve(min_t_hat*m+b==100*fut_min_t, max_t_hat_p*m+b==100*past_max_t);
         temp_hat(max_index_p:min_index)= solm*temp_hat(max_index_p:min_index)+solb;

         [solm, solb] = solve(min_t_hat*m+b==100*fut_min_t, max_t_hat*m+b==100*fut_max_t);

             temp_hat(min_index:max_index)= solm.*temp_hat(min_index:max_index)+solb;

         %Error

         e(max_index_p:max_index)=temp_hat(max_index_p:max_index)-0.1*Temp_actual(max_index_p:max_index)';
         e1(max_index_p:max_index)=0.01.*e(max_index_p:max_index);
     else
         [solm, solb] = solve(max_t_hat*m+b==100*fut_max_t, min_t_hat_p*m+b==100*past_min_t);
         temp_hat(min_index_p:max_index)= solm*temp_hat(min_index_p:max_index)+solb;

         [solm, solb] = solve(min_t_hat*m+b==100*fut_min_t, max_t_hat*m+b==100*fut_max_t);
         temp_hat(max_index:min_index)= solm*temp_hat(max_index:min_index)+solb;

         %Error

         e(min_index_p:min_index)=(temp_hat(min_index_p:min_index)-0.1*Temp_actual(min_index_p:min_index)');
         e1(min_index_p:min_index)=0.01*e(min_index_p:min_index);
     end
     %Weight Update
     w1=w1-alpha1*phi1*(w1'*phi1+k_v*B*e1(i-23:i)')';
     w2=w2-alpha2*phi2*(w2'*phi2+k_v*B*e1(i-23:i)')';
     w3=w3+alpha3*phi3*error(i:i+23);
     min_t_hat_p=min_t_hat;   
      max_t_hat_p=max_t_hat;
     %Temperature for next day calculation
     tempn24=temp24;

     flag=flag+1;
  end
plot(temp_hat);
hold on;
plot(Temp_actual,'r');
  end

If some body can help me diagnose the problem it will be great.

[Star Strider]

The paper is behind the IEEE paywall. I cannot access it.

[Greg Heath]

Why can't you use the NN Toolbox?

[Siddharth Singh]

Thanks for the reply... My professor asked not to use. I found out the problem it was the output scaling. Thanks again!