Assignment PTE586 3
Zhefeng, Zhang
Sep, 15 2010
1.Explain how would you design and train an artificial neural network to forecast the
Gas rate
Because of the known values, an ANN can be designed in the following manner: its
consist of 3 layers, the input layer, the hidden layer and the output layer. There are 4
neurons (gas flow rate, water flow rate, well head pressure, well head temperature) in
the input layer, and 2 in the hidden layer, 1 in the output layer (gas rate). The structure
is 421.
Using 70% of the data for training ANN, the left 30% for testing the training result.
Pre-input: using the non-linear sigmoid function f(x)=1/(1 + e- x), if the data does not
fall into [0,1]
Using x = (x1 - x1min)/(x1max - x1min)
From the input layer node X to hidden layer node:
The equation for input layer to hidden layer is
hni1 = Mhn11 x1 + Mhn12 x2 + Mhn13 x3 + Mhn14x4
hni2 = Mhn21 x1 + Mhn22 x2 + Mhn23 x3 + Mhn24x4
In which, Mhn is the hidden layer nodes matrix, 2 lines 4 columns. Hni is the input
data to hidden layer nodes.
The equation for output data from the hidden layer node:
hno1 = f(hni1)=1/(1+exp(- hni1))
hno2 = f(hni2)=1/(1+exp(- hni2))
From the hidden layer to the output layer
oni1 = Mon1,1hno1 +Mon1,2 hno2 + Mon1,3hno3 +Mon1,4hno4
In which, Mon is the output node matrix, 1 line, 2 columns. Oni is the output data from
output layer.
2. Present a graph of the ANN and show clearly what are the inputs and outputs of
ANN the
3. What conventional statistical method is needed to use before applying the ANN
Before applying ANN, conventional extrapolate method is needed, since the data from
8.am to 10.am and Sunday is missing. The data can be extrapolated from the data prior
to 8am and after 10am. The average of 2 hours prior to 8am and after 10am can work as
a substitute data. And the missing Sunday data can be extrapolated and substituted by
the average of Saturday and Friday whole days data.
4. What other soft computing should be used in conjunction with ANN and
incorporate the qualifiers in the solution.
Fuzzy logic should be used along side the ANN.
The fuzzy qualifier much higher is also one of the fuzzy sets represents the output of the
prior daily average rate. Define 5 fuzzy sets on X as much lower, lower, equal, higher,
much higher. These sets are drawn into curve function in X. Let Y represents the
methods of preprocessing the data. If X is much higher, the method is calculating the
daily average rate by daily water flow deducted by 24 hours.
When the hourly temperature is near or above the top quartile, Fuzzy sets should be set
up as below the lower quartile, near the 2nd quartile, near the 3rd quartile, and near or
above the top quartile, using trapezoidal membership functions. Then near or above the
top represents a 2 to 3 times growth in gfr. Also adding the limit that the temperature
does not exceed the boiling point temperature.