Borrar filtros
Borrar filtros

How evaluate multi-step forecasting performance for large dataset?

1 visualización (últimos 30 días)
How evaluate multi-step forecasting performance for large dataset?
What I means is that to use NARX for forecasting I have to:
1. create closed loop network with
netc = closeloop(net);
2. construct target prediction data with
targetSeriesPred = [targetSeries(end-delay+1:end), con2seq(nan(1,N))], i.e. use a number of nan equal to forecasting horizon
3. prepare data with
[Xs,Xi,Ai,Ts] = preparets(netc,inputSeriesPred,{},targetSeriesPred);
4. get simulation/prediction results with
yPred = netc(Xs,Xi,Ai);
5. evaluate forecasting performance (regarding olny one step-ahead prediction) with
perf = perform(net,yPred,targetSeriesVal);
Now suppose I have a large dataset of data concerning 3 years of hourly sampled data, I can utilize only the first year for traning, validation and test phase ('divideblock' of course). I don't know how to get a unique performance value for the remaining data, 2 years, for a forecasting horizon for example of only 6 samples. I hope I to have been clear enough. Thanks in advance for your replies/suggestions.

Respuesta aceptada

Greg Heath
Greg Heath el 14 de Abr. de 2014
Editada: Greg Heath el 25 de Mzo. de 2016
How evaluate multi-step forecasting performance for large dataset? Asked by EanX 14Apr2014
A. Design an openloop net
1. Use training data and nncorr or fft to estimate
a. Significant autocorrelation output feedback lags
b. Significant input/output crosscorrelation input lags
2. Use trial and error to determine number of hidden layer nodes
3. Use divideblock in narxnet to obtain multiple designs
4. Choose the best design based on validation set error.
5. Use the test set on the best design to obtain an UNBIASED
estimate of performance
B. Convert to a closeloop design
1. [ netc Xci Aci ]= closeloop(neto, Xoi, Aoi)
2. Use preparets and test netc on the original data
3. If performance is unsatisfactory, train netc intialized with
the initial conditions Xoi, Aoi and final weights of neto.
4. To predict beyond the original data, use the data at the end of
the test set to intialize the delay buffer and a new external input.
5. Use preparets and netc to predict beyond the test data.
  2 comentarios
EanX el 15 de Abr. de 2014
Thanks Greg, but what I mean is that for evaluate performance for, example, six steps ahead I can:
targetSeriesPred=[targetSeries(end-delay+1:end), con2seq(nan(1,steps_ahead))];
fprintf('Performance closed loop (%d step ahead):%d\n',steps_ahead,perfCloseLooop);
e= targets-outputs;
But this is only for six steps ahead, to evaluate a unique MSE, taking count of net's performance for six steps ahead in closed loop, not only one step ahead, for 2 years of data starting from the first used for traning validation and test, I suppose that I could iterate the above code in order to construct a global "outputs" and then avaluate error and MSE, but this require retraining the NARX at each step? This is correct?
Greg Heath
Greg Heath el 21 de Abr. de 2014
Sorry, I don't understand.
1. If this is a single series you should use narnet, not narxnet.
2. Are the sample summary statistics (mean, variance, significant correlation coefficients and corresponding lags) time-independent?
3. What are the significant correlation lags?
4. How are you designing netc in the first place?
5. What ratio of trn/val/tst are you planning to use? I doubt if Ntst > Ntrn is a reasonable goal.
6. Try practicing on a MATLAB example dataset
help nndatasets
doc nndatasets
7. See

Iniciar sesión para comentar.

Más respuestas (0)


Más información sobre Sequence and Numeric Feature Data Workflows en Help Center y File Exchange.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by