Forecasting electricity load based on various parameters using neural networks.

I have a task of predicting the electricity load for 2 months of a region based on various parameters, like avg. rainfall, avg. solar radiation, winter or summer. etc.

There are two sets of data given. 1) To analyse and form a model - This contains the parameter and the electricity load values too which should be used to test equation and create a model that will be employed to do the actual work 2) The set of data based on which the prediction is to be made.

I am using neural net fitting to make the model. More clearly the Bayesian Regularization. The train models with this gives me 3.8% Mean avg. percentage error. But the actual forecasting model gives a very high 28% Mean avg. percentage error.

My question is

1) How can I better modify the data to get a better result. I am now thinking to use the Smoothing Spline function to remove the unevenness in the test data and then use the neural net fitting.

2) Is there a better mathematical function that I can use to tackle this problem ?

Answer 1

[Greg Heath]

 How may inputs? How many outputs? How much data?

 Typical time between measurements?

 Are measurements equally spaced ?

 How are training and testing subsets chosen? Size ratio?

 *How different are the summary statistics of the training and testing subsets?*

 How are all the matrices normalized?

 Why fitting instead of timeseries?

 Why TRAINBR instead of TRAINLM?

 What fraction of the average target variance is not modelled? 

     NMSE = mse(target-output) / mean(var(target',1))

How many models did you design before choosing this one?

Hope this helps.

Greg

[sangam singh]

I have 8 inputs:

• Avg. daily temperature
• Avg. daily cloud cover
• Avg. daily wind speed
• Avg. daily humidity
• Avg. daily precipitation
• Avg. daily solar radiation
• Day Type (Weekend, Weekday, Other Holidays)
• School Holiday (0 or 1)
• Winter (0 or 1)

And I have one output, that is the Electricity Load Consumption.

All the data are per day values.

I am using TRAINBR because it gave me lower Mean Average Percentage Error.

The data set for training is for 2 years.

Further Note:

I have not yet considered time series. And what do you mean by are all the matrices normalized ?

Before using the final model, I created several model by using TRAINBR with various numbers of Neurons.

[Greg Heath]

Are the inputs and targets raw data or are they transformed or normalized ?

The better question is:

Do training and testing data have comparable summary statistics?

Before you attempt time series modeling obtain the target autocorrelation and input/target cross correlation functions. Obtaining the delays of the significant values will help estimating what delays to use in the prediction functions.

It may be useful to look up my timeseries posts and tutorials in the NEWSGROuP and ANSWERS.

ope this helps.

Greg