2 dimensional features for GaussianHMM prediction

[Cheng Chen]

Hello all:

I go through the code from FHMM disaggregation algorithm and have a question about it. The FHMM algorithm in the NILMTK package is only using 1-Dimensional data, which is the active power, for the train and predict process. I was wondering that what if we use 2-Dimensional of data (which is active & reactive power) to train and predict? Theoretically, it would make our prediction much more reliable. So I go through the GaussianHMM module, which is the critical one for the FHMM disaggregation, and find no where to input a 2-Dimensional training data.

I have found that the GaussianHMM module has an attribute called "n_features" which explained as the "Dimensionality of the Gaussian emissions". But there's no parameter in the module to change that attribute.

For example, I have a dataframe of the active power(W) and reactive power(VAR) of a fridge. I tried this:

In[1]:    data.head()

Out[1]: 

                                      W      VAR

2013-06-07 00:00:00  0.111  2.483

2013-06-07 00:00:01  0.200  2.547

2013-06-07 00:00:02  0.152  2.480

2013-06-07 00:00:03  0.159  2.444

2013-06-07 00:00:04  0.215  2.510

In[2]:    data.shape

Out[2]: (2000, 2)

In[3]:    active = np.array(data.W.tolist())

    ...:    reactive = np.array(data.VAR.tolist())

    ...:    total = np.concatenate((active, reactive))

    ...:    length = [len(active), len(reactive)]

In[4]:    model = hmm.GaussianHMM(n_components=2)

    ...:    model.fit(total.reshape(-1,1), length)

Out[4]: 

GaussianHMM(algorithm='viterbi', covariance_type='diag', covars_prior=0.01,

      covars_weight=1, init_params='stmc', means_prior=0, means_weight=0,

      min_covar=0.001, n_components=2, n_iter=10, params='stmc',

      random_state=None, startprob_prior=1.0, tol=0.01, transmat_prior=1.0,

      verbose=False)

In[5]:    pred=model.predict(total.reshape(-1,1), length)

Out[5]: array([1, 1, 1, ..., 1, 1, 1])

In[6]:    len(pred)

Out[6]: 4000

The "pred" supposed to return a list of states of only 2000 elements. But it returns 4000, which means it regard my "total" as a 1-Dimensional data to disaggregate.

So anyone have ideas about using 2-Dimensional data to train and predict the states?