CN110119816A - A kind of load characteristic self-learning method suitable for non-intrusion type electric power monitoring - Google Patents

Abstract

The invention provides a load characteristic self-learning method suitable for non-intrusive power monitoring, comprising: acquiring a load data sequence after the occurrence of a load event as an input sample; randomly adding noise to the sample; determining noise reduction according to the length of the data sample sequence The number of neurons in the input layer of the autoencoder to generate the input layer and the output layer; determine the number of neurons in the hidden layer of the autoencoder and generate the hidden layer; set the training error limit of the denoising autoencoder; initialize the denoising autoencoder Inter-layer mapping parameters; calculate the reconstruction error of the sequence to the input sequence according to the mapping parameters; if the reconstruction error is less than the training error limit, the hidden layer node value is extracted as the abstract feature of the load event, and if it is greater than the training error limit, the gradient descent algorithm is used Update the mapping parameters between the input layer and the hidden layer, and the hidden layer and the output layer. The invention realizes the compressed sensing of the data sequence, thus realizes the learning of abstract features, and provides a global interpretation of the load event data curve.

Description

A self-learning method of load characteristics suitable for non-intrusive power monitoring

technical field

The invention belongs to the technical field of power systems, and relates to a load characteristic self-learning method suitable for non-intrusive power monitoring.

Background technique

The non-invasive load monitoring (NILM) technology includes four basic contents: 1) data and preprocessing acquisition; 2) event detection; 3) feature extraction; 4) load identification. The principle of which together constitute the non-intrusive load monitoring system is shown in Figure 1. When the system is working, the data acquisition and preprocessing module first collects and calculates the total load data (active power, reactive power, voltage, current, etc.), and transmits it to the event detection module; the event detection module can detect when it happened. Load events (load input or removal); the feature extraction module extracts load event features (including steady-state features and transient features) after the occurrence of load events according to the results of event detection; finally, the load identification module extracts load event features according to the extracted load event features, Through the classification and identification algorithm, the load events are classified and identified. Among them, the load feature extraction module plays an important role in NILM. Only when correct and effective load features are extracted can these features be further used to identify the load through the load classification and identification algorithm.

The current research on load characteristics mainly focuses on the steady-state and transient physical characteristics after load switching events, including: active power, active power, reactive power, current, voltage and their differences, current-voltage trajectories, and higher harmonics. wave characteristics, etc. These extracted feature quantities all have clear physical meanings, which need to be set manually during feature extraction, and then obtained by calculating the power data collected by the data acquisition module. When calculating these features, it is often only based on local data points. Taking the peak power value of the feature quantity as an example, it only uses one data point after the occurrence of the load event, and the feature of active power difference is only used at a certain moment. Two data points before and after. The local data points have a certain interpretability for the corresponding load event data curve, which can reflect the general characteristics of the curve, but still lacks the global interpretability of the load event data curve.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention proposes a load feature self-learning method suitable for non-intrusive power monitoring. It is not necessary to manually set which specific physical features need to be extracted by the feature extraction module, but to learn independently can reflect the nature of the load event. The abstract feature of the feature, the data source of feature learning in this method is the load data sequence corresponding to the load event time calibrated by the switching event module.

In order to achieve the above object, the present invention provides the following technical solutions:

A self-learning method for load characteristics suitable for non-intrusive power monitoring, comprising the following steps:

Step 1. Obtain the load data sequence after the occurrence of the load event as an input sample;

Step 2: Randomly add noise to the input sample;

Step 3: Determine the number of neurons in the input layer of the noise reduction autoencoder according to the length of the data sample sequence, and generate an input layer and an output layer;

Step 4: Determine the number of neurons in the hidden layer of the noise reduction autoencoder, and generate the hidden layer;

Step 5. Set the training error limit of the noise reduction autoencoder;

Step 6: Initialize the input layer and the hidden layer of the noise reduction autoencoder, and the mapping parameters between the hidden layer and the output layer, and the parameters include weights and biases;

Step 7: Calculate the reconstruction error of the output sequence for the input sequence according to the mapping parameters between the input data sequence and each layer;

Step 8: Judging whether the reconstruction error is less than the set training error limit, if the reconstruction error is less than the training error limit, go to step ten, and if the reconstruction error is greater than the training error limit, go to step nine;

Step 9. Use the gradient descent algorithm to update the input layer and the hidden layer, and the mapping parameters between the hidden layer and the output layer;

Step 10: Extract the node value of the hidden layer as the abstract feature of the load event.

Further, in the third step, the structure of the output layer is the same as that of the input layer, and the number of neurons in the output layer is the same as that of the input layer.

Further, in the fourth step, the number of neurons in the hidden layer is less than the rated number of neurons in the input layer and the output layer.

Further, in the step 6, the mapping function between the input layer and the hidden layer is defined as:

Y=f _θ (X')=S(WX'+b) (1)

In formula (1), S(X) is the activation function of the noise reduction autoencoder, θ is the encoding parameter, which is composed of the weight W and the bias b;

The mapping function between the hidden layer and the output layer is defined as:

Z=f _θ' (Y)=S(W'Y+b') (2)

In formula (2), θ' is a decoding parameter, which is composed of weight W' and bias b'.

Further, in the step 7, the calculation formula of the reconstruction error is:

where l is the number of neurons in the input layer of the autoencoder.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

The invention utilizes the noise reduction self-encoder model to encode and then decode the input load data sequence, so as to realize the compressed sensing of the data sequence, thereby realizing the learning of abstract features. The data source of the feature self-learning in the method of the present invention is the load data sequence corresponding to the load event time calibrated by the switching event module, and the global interpretation of the load event data curve is realized.

Description of drawings

Figure 1 is a basic schematic diagram of a non-intrusive load monitoring system.

Figure 2 is a structural diagram of an autoencoder.

FIG. 3 is a flowchart of a load characteristic self-learning method suitable for non-intrusive power monitoring.

Detailed ways

The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

The present invention is realized by using a noise reduction auto-encoder model. The auto-encoder is a special neural network, that is, the output is the same as the input, and the model adjusts parameters through training, so that the input can recover the original input as much as possible by means of feature encoding and decoding. These values after feature coding and transformation are the abstract features representing the input signal, and the general autoencoder structure is shown in Figure 2.

Taking the time series corresponding to different load switching events as the input of the autoencoder, taking a specific load switching event sample as an example, assuming the corresponding number of samples is k, the sample set is x={x ⁽¹⁾ ,x ⁽²⁾ ...x ^(k) }, any sample x ⁽ⁱ⁾ is a time series of length l, that is, x ⁽ⁱ⁾ is an l-dimensional vector, and the number of neurons in the input layer of the self-encoder is designed is l, and the number of neurons in the middle hidden layer is designed to be m. Since the autoencoder uses the back-propagation algorithm to optimize the reconstruction error of the input data, even if the target output y ⁽ⁱ⁾ → x ⁽ⁱ⁾ , the neural network is forced to learn A compressed representation of the input data, that is, x ⁽ⁱ⁾ must be reconstructed from the m-dimensional hidden neuron activation vector α ⁽ⁱ⁾ ∈ R ^m . If any sample in the sample set is completely random, for example, each input x ⁽ⁱ⁾ is an independent and identically distributed Gaussian random variable that is completely independent of other input variables, this learning process will be difficult to carry out, but if the input The sample data implies some specific structure, so this algorithm can find the correlation between the input sample data. After network training, the hidden layer activation vector α ^{(i ) corresponding to each input sample x (i} ) is equivalent to the abstract feature vector after dimension reduction (learning). By adding some random noise to the input of the autoencoder, the autoencoder will be able to obtain the ability to extract abstract features from the disturbed input data, and the robustness of the autoencoder will be enhanced.

A self-learning method of load characteristics suitable for non-intrusive power monitoring proposed by the present invention, the process of which is shown in Figure 3, including the following steps:

Step 1. Obtain the load data sequence after the occurrence of the load event as an input sample:

Taking the active power data as an example, the starting time of the load event and the corresponding steady-state and transient processes are calibrated by the event detection algorithm, and the active power data sequence from the starting time to the steady-state process is used as the input of the self-encoder. sample.

Step 2. Randomly add noise to the input samples:

The purpose of this step is to turn the input sample X into a noisy sample X' to simulate the effect of random disturbances on the feature learning ability of the autoencoder. If there is a small reconstruction error, the robustness of its feature learning ability is considered to be enhanced. An autoencoder that artificially adds noise is called a noise reduction autoencoder.

Step 3: Determine the number of neurons in the input layer of the denoising autoencoder according to the length of the data sample sequence, and generate the input layer and output layer:

If the sequence length of the input sample X' after adding noise is 1, the number of neurons in the input layer is also set to 1, that is, there is a one-to-one mapping relationship between the input sample sequence and the neurons in the input layer. According to the principle of autoencoder feature learning, the reconstruction error of the input data sequence needs to be minimized, so the output layer should maintain the same structure as the input layer, that is, the number of neurons in the output layer is the same as the input layer.

Step 4. Determine the number of hidden layer neurons of the denoising autoencoder and generate the hidden layer:

In step 3, it has been determined that the rated number of neurons in the input layer and the output layer is l. The selection of the number of neurons in the hidden layer k should follow the principle of k<l. This is to satisfy the compression of high-dimensional input vectors, compression into a lower-dimensional abstract feature vector, so as to realize the compression and extraction of data features.

Step 5: Set the training error limit of the noise reduction autoencoder.

The training error bound can be thought of as an upper bound on the acceptable reconstruction error.

Step 6: Initialize the input layer and the hidden layer of the noise reduction autoencoder, and the mapping parameters between the hidden layer and the output layer, and the parameters include weights and biases.

The mapping function between the input layer and the hidden layer can be defined as:

Y=f _θ (X')=S(WX'+b) (1)

In formula (1), S(X) is the activation function of the noise reduction autoencoder, θ is the encoding parameter, which is composed of the weight W and the bias b;

The mapping function between the hidden layer and the output layer can be defined as:

Z=f _θ' (Y)=S(W'Y+b') (2)

In formula (2), θ' is a decoding parameter, which is composed of weight W' and bias b';

Step 7: Calculate the reconstruction error of the output sequence with respect to the input sequence according to the mapping parameters between the input data sequence and each layer.

The formula for calculating the reconstruction error is:

Note that in equation (3), the output sequence Z is calculated on the basis of the input sequence X' that has been subjected to noise processing, but the original input sequence X is still involved in the calculation of the reconstruction error.

Step 8. Judge whether the reconstruction error is less than the set training error limit. If the reconstruction error is less than the training error limit, it is considered that the noise reduction autoencoder model has learned the abstract features that have good interpretability for the input data, and go to step ten.

If the reconstruction error is greater than the training error limit, it is considered that the denoising autoencoder model has not learned the abstract features that have good interpretability for the input data, and go to step 9.

Step 9. Use the gradient descent algorithm to update the input layer and the hidden layer, and the mapping parameters between the hidden layer and the output layer.

Step 10: Extract hidden layer node values as abstract features of load events.

The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be pointed out that for those skilled in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims (5)

1. a load characteristic self-learning method applicable to non-intrusive power monitoring, is characterized in that, comprises the steps: Step 1. Obtain the load data sequence after the occurrence of the load event as an input sample; Step 2: Randomly add noise to the input sample; Step 3: Determine the number of neurons in the input layer of the noise reduction autoencoder according to the length of the data sample sequence, and generate an input layer and an output layer; Step 4: Determine the number of neurons in the hidden layer of the noise reduction autoencoder, and generate the hidden layer; Step 5. Set the training error limit of the noise reduction autoencoder; Step 6: Initialize the input layer and the hidden layer of the noise reduction autoencoder, and the mapping parameters between the hidden layer and the output layer, and the parameters include weights and biases; Step 7: Calculate the reconstruction error of the output sequence for the input sequence according to the mapping parameters between the input data sequence and each layer; Step 8: Judging whether the reconstruction error is less than the set training error limit, if the reconstruction error is less than the training error limit, go to step ten, and if the reconstruction error is greater than the training error limit, go to step nine; Step 9. Use the gradient descent algorithm to update the input layer and the hidden layer, and the mapping parameters between the hidden layer and the output layer; Step 10: Extract the node value of the hidden layer as the abstract feature of the load event. 2. The load characteristic self-learning method suitable for non-intrusive power monitoring according to claim 1, characterized in that: in the step 3, the output layer and the input layer have the same structure, and the number of neurons in the output layer is the same as that in the input layer. same. 3. The load characteristic self-learning method suitable for non-invasive power monitoring according to claim 1, wherein in the step 4, the number of neurons in the hidden layer is less than the number of rated neurons in the input layer and the output layer . 4. The load characteristic self-learning method suitable for non-intrusive power monitoring according to claim 1, wherein in the step 6, the mapping function between the input layer and the hidden layer is defined as: Y=f _θ (X')=S(WX'+b) (1) In formula (1), S(X) is the activation function of the noise reduction autoencoder, θ is the encoding parameter, which is composed of the weight W and the bias b; The mapping function between the hidden layer and the output layer is defined as: Z=f _θ' (Y)=S(W'Y+b') (2) In formula (2), θ' is a decoding parameter, which is composed of weight W' and bias b'. 5 . The load characteristic self-learning method suitable for non-intrusive power monitoring according to claim 1 , wherein in the step 7, the calculation formula of the reconstruction error is: 5 . where l is the number of neurons in the input layer of the autoencoder.