Energy data context information feature fusion method based on reinforcement learning

Abstract

The present invention relates to the field of data fusion technology, and in particular to a method for fusion of energy data context information features based on reinforcement learning, the method specifically comprising: constructing an energy peak connection intensity index according to the degree of association between energy data, using an ARIMA prediction algorithm to predict energy data sequences to obtain prediction feature sequences, constructing prediction feature correlation coefficients according to the quality of the predicted data in the prediction feature sequences and the association with the original data, and then dividing data sequences that truly have strong associations, and then using deep learning to fuse context information features. Thus, the fusion of energy data context information features is achieved, and the problem that the traditional feature fusion algorithm only considers the association but not the essence of the association, resulting in too much associated data, and ultimately resulting in low accuracy of features after feature fusion is solved, and the accuracy of energy data context feature information fusion is improved.

Description

Technical Field

[0001] The present application relates to the technical field of data fusion, and in particular to a method for fusing contextual information features of energy data based on reinforcement learning. Background Art

[0002] With the advancement of digitalization, the digital content of various industries is gradually enriched. Among them, the energy industry involves a lot of important data, so the relevant processing of energy data has gradually become an important part of the research on data processing methods. Energy data usually contains a lot of sensitive data related to the enterprise, such as the enterprise's equipment energy consumption information, equipment operation data, user identity information and password information of the energy platform, etc. Once such data is leaked, it may cause huge losses, so sensitive data in energy data is usually protected.

[0003] However, energy data involves a large amount of data, and there are many different degrees of correlation between the data. Therefore, it is easy to cause inaccurate identification in the process of sensitive data identification. Therefore, before identifying energy sensitive data, the context information feature fusion method based on reinforcement learning is usually used to fuse the information features of energy data to reduce the difficulty of identifying sensitive data later. The traditional feature fusion method extracts features from all data information and then fuses them. However, the context information of energy data involves a wide range of data information, and there are many related data information. If feature fusion is performed directly, more data features that have no substantial correlation will be fused, resulting in inaccurate data features after fusion, which in turn affects the accuracy of identifying energy sensitive data. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an energy data context information feature fusion method based on reinforcement learning to solve the existing problems.

[0005] The energy data context information feature fusion method based on reinforcement learning of the present invention adopts the following technical solutions:

[0006] An embodiment of the present invention provides a method for fusing energy data context information features based on reinforcement learning, the method comprising the following steps:

[0007] Collect each energy data sequence in each type of energy data;

[0008] The polynomial curve of each energy data sequence is obtained by a polynomial curve fitting algorithm, and the super-standard and sub-standard peak point data sequences of each energy data sequence are obtained according to the peak point size and corresponding time in the polynomial curve; in the same type of energy data, the energy peak connection intensity index of each energy data sequence is obtained according to the data changes in the super-standard and sub-standard peak point data sequences; the energy peak connection intensity index of all energy data sequences is clustered by a clustering algorithm to obtain a strongly connected energy data sequence, and the prediction feature sequence of each strongly connected energy data sequence is obtained by combining the ARIMA prediction algorithm; the prediction reliability weight of each data in the prediction feature sequence is obtained according to the data changes in the strongly connected energy data sequence and the prediction feature sequence; the prediction excellence index of each data in the prediction feature sequence is constructed according to the prediction reliability weight and the autocorrelation changes before and after the data in the prediction feature sequence is added to the strongly connected energy data sequence; the prediction feature correlation coefficient of each prediction feature sequence is obtained according to the prediction excellence index and the data difference in different prediction feature sequences;

[0009] According to the prediction feature correlation coefficient combined with the clustering algorithm, each highly correlated energy data sequence is obtained, and the context feature fusion result of each highly correlated energy data sequence is obtained through the reinforcement learning strategy feature fusion method.

[0010] Preferably, the polynomial curve of each energy data sequence is obtained by a polynomial curve fitting algorithm, and each super-standard and sub-standard peak point data sequence of each energy data sequence is obtained according to the peak point size and corresponding time in the polynomial curve, specifically:

[0011] For each energy data sequence, a polynomial curve of the energy data sequence is obtained by fitting the energy data sequence through a polynomial curve fitting algorithm; the mean of all peak point data in the polynomial curve is calculated; the peak point with a value greater than the mean is regarded as an over-standard peak point, and the peak point with a value less than the mean is regarded as a low-standard peak point;

[0012] All the super-standard peak points on the polynomial curve are used as the input of the regional growing algorithm, where each super-standard peak point is used as the initial growth point, and the algorithm outputs each growth area; the sequence composed of all the super-standard peak points in each growth area is used as the super-standard peak point data sequence of the energy data sequence; and each low-standard peak point data sequence is obtained by obtaining each super-standard peak point data sequence.

[0013] Preferably, the energy peak connection intensity index of each energy data sequence is obtained according to the data changes in the super-standard and sub-standard peak point data sequences, specifically:

[0014] The characteristic disturbance index of each energy data series is obtained according to the numerical difference and length change in each over-standard and under-standard peak data series;

[0015] Calculate the cosine similarity between the b-th energy data sequence and each energy data sequence; calculate the sum of all the cosine similarities of the b-th energy data sequence; calculate the absolute value of the difference between the characteristic susceptibility index of the b-th energy data sequence and each energy data sequence; calculate the sum of all the absolute values ​​of the differences of the b-th energy data sequence; calculate the sum of the absolute values ​​of the differences and the sum of the preset parameters, recorded as the first sum; and use the ratio of the sum of the cosine similarities to the first sum as the energy peak connection intensity index of the b-th energy data sequence.

[0016] Preferably, the characteristic disturbance susceptibility index of each energy data sequence is obtained according to the difference and length change between each super-standard and sub-standard peak point data sequence, and the expression is:

[0017]

[0018]

[0019] In the formula, C b is the characteristic disturbance index of the b-th energy data sequence, D b is the significance index of exceeding the standard for the bth energy data series, F and E are the total number of low-standard peak data series and exceeding-standard peak data series of the bth energy data series, respectively. are the means of all data in the i-th and k-th low standard peak data sequences of the b-th energy data sequence, are the ith and kth low standard peak data sequences of the bth energy data sequence, respectively. len(*) is the function for obtaining the length of the data sequence. are the lengths of the i-th and k-th low standard peak data sequences of the b-th energy data sequence, is the mean of all data in the jth super-standard peak point data sequence of the bth energy data sequence, is the jth super-standard peak point data sequence of the bth energy data sequence, It is the mean length of all low-standard peak point data sequences and all super-standard peak point data sequences of the b-th energy data sequence.

[0020] Preferably, the energy peak connection intensity index of all energy data sequences is clustered by a clustering algorithm to obtain a strongly connected energy data sequence, and the prediction feature sequence of each strongly connected energy data sequence is obtained by combining the ARIMA prediction algorithm, specifically:

[0021] The energy peak connection intensity index of all energy data sequences is clustered by clustering algorithm to obtain two clusters; the mean of all elements in each cluster is calculated; the cluster with the largest mean is taken as the strong connection cluster; each energy data sequence in the strong connection cluster is taken as the strong connection energy data sequence;

[0022] The ARIMA prediction algorithm is used to predict the data of each strongly connected energy data series for a preset number of moments, and the sequence composed of all the predicted data of each strongly connected energy data series is used as the prediction feature sequence of each strongly connected energy data series.

[0023] Preferably, the step of obtaining the prediction reliability weight of each data in the prediction feature sequence according to the data changes in the strongly connected energy data sequence and the prediction feature sequence specifically includes:

[0024] Calculate the mean of all data in each strongly connected energy data sequence; calculate the absolute value of the difference between each predicted data in the predicted feature sequence of each strongly connected energy data sequence and the mean; calculate the calculation result of an exponential function with a natural constant as the base and the opposite number of the absolute value of the difference as the exponent; and use the ratio of the calculation result to the corresponding collection time of each predicted data as the prediction reliability weight of each predicted data.

[0025] Preferably, the construction of the prediction excellence index of each data in the prediction feature sequence specifically includes:

[0026] Calculate the autocorrelation coefficient ρ of the b′th strongly connected energy data sequence n ; Calculate the autocorrelation coefficient ρ of the energy data sequence at the mth prediction moment of the b′th strongly connected energy data sequence n+m ; The specific expression for calculating the prediction excellence index of each data in the prediction feature sequence is:

[0027]

[0028] In the formula, is the prediction excellence index of the mth prediction data in the b′th prediction feature sequence, is the prediction reliability weight of the mth prediction data in the b′th prediction feature sequence, X is the number of prediction data in the b′th prediction feature sequence, N m and N l are respectively the data values ​​of the mth and lth prediction data in the b′th prediction feature sequence, and γ is the preset first parameter.

[0029] Preferably, the energy data sequence at the prediction moment is specifically:

[0030] The 1st to yth predicted data of the strongly connected energy data sequence are added to the strongly connected energy data sequence, and the formed new sequence is used as the yth predicted moment energy data sequence of the strongly connected energy data sequence.

[0031] Preferably, the prediction feature correlation coefficient of each prediction feature sequence is obtained according to the prediction excellence index and the data difference in different prediction feature sequences, specifically including:

[0032] The ApEn value of each prediction feature sequence is obtained by the ApEn algorithm; the specific expression for calculating the prediction feature association coefficient of each prediction feature sequence is:

[0033]

[0034] Where, L b′ is the prediction feature correlation coefficient of the b′th prediction feature sequence, e * is an exponential function with the natural constant e as the base, u is the total number of strongly connected energy data sequences, X is the number of predicted data in the b′th predicted feature sequence, are the prediction excellence indexes of the mth prediction data in the b′th and g′th prediction feature sequences, respectively. To preset the second parameter, ApEn b′ ,ApEn g′ are the ApEn values ​​of the b′th and g′th predicted feature sequences.

[0035] Preferably, the highly correlated energy data sequences are obtained by combining the predicted feature correlation coefficient with a clustering algorithm, and the context feature fusion results of the highly correlated energy data sequences are obtained by a reinforcement learning strategy feature fusion method, specifically:

[0036] Perform K-mediods clustering on the predicted feature correlation coefficients of all predicted feature sequences to obtain two clusters, and take the cluster with the largest mean of all predicted feature correlation coefficients in the cluster as the correlation cluster;

[0037] The energy data sequences corresponding to all predicted feature sequences in the associated clustering cluster are recorded as high-correlation energy data sequences; each high-correlation energy data sequence is used as the input of the autoencoder, and the output is the sequence features of each high-correlation energy data sequence; the sequence features of the b″th high-correlation energy data sequence and the sequence features of other high-correlation energy data sequences are used as input, and the feature fusion method of the reinforcement learning strategy is adopted, and the output is the feature fusion result.

[0038] The present invention has at least the following beneficial effects:

[0039] The present invention obtains a feature disturbance susceptibility index and an over-standard significance index by analyzing the degree of over-standard of energy data peak and the susceptibility of data to interference, constructs an energy peak connection strength index, and preliminarily eliminates a portion of data with weak correlation; predicts the energy data sequence by using an ARIMA prediction algorithm to obtain a prediction feature sequence; constructs a prediction feature correlation coefficient by analyzing the prediction effect of the data in the prediction feature sequence and the degree of correlation with the original data, further determines the data with strong correlation with the energy data sequence currently required to be analyzed, reduces the amount of data when performing context information feature fusion, and then divides the data sequence that truly has a strong correlation, and then uses deep learning to perform context information feature fusion; solves the problem that the traditional feature fusion algorithm only considers the correlation but not the essence of the correlation, resulting in excessive correlation data, and ultimately resulting in low accuracy of the feature after feature fusion, thereby improving the accuracy of energy data context feature information fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 A flow chart of the energy data context information feature fusion method based on reinforcement learning provided by the present invention;

[0042] Figure 2 Schematic diagram of the steps of the energy data context information feature fusion method based on reinforcement learning. DETAILED DESCRIPTION

[0043] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the energy data context information feature fusion method based on reinforcement learning proposed by the present invention, its specific implementation method, structure, features and effects are described in detail as follows in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0044] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0045] The specific scheme of the energy data context information feature fusion method based on reinforcement learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0046] An embodiment of the present invention provides a method for fusing energy data context information features based on reinforcement learning.

[0047] Specifically, the following energy data context information feature fusion method based on reinforcement learning is provided, please refer to Figure 1 , the method comprises the following steps:

[0048] Step S001, collecting energy data from the energy data platform and performing preprocessing.

[0049] In the database of the energy data platform, the energy data stored in the energy data platform in the past month is obtained and classified according to the existing categories in the energy data platform. All energy data in each data of each category of energy data are arranged in ascending order according to their time sequence, and each energy data sequence of each category of energy data is constructed, where the bth energy data sequence of the ath category of energy data is used express.

[0050] Since data may be lost during the transmission process, this embodiment uses existing data and a mean filling method to fill the lost data. The mean filling algorithm is a well-known technology, and this embodiment will not repeat its process.

[0051] Step S002, constructing an energy peak connection intensity index according to the degree of correlation between energy data, using the ARIMA prediction algorithm to predict the energy data sequence to obtain a prediction feature sequence, and constructing a prediction feature correlation coefficient according to the quality of the predicted data and its correlation with the original data.

[0052] Take the energy data sequences of the a-type energy data as an example. There are many categories of data in energy data, and these data may have certain correlations. For an energy data sequence, its context information features may include its own temporal context features, context features of the order of storage (such as when the last stored sequence is power data, and the next stored sequence is also power data, then the data sequence may also be power data) and context features of data correlation, etc. However, since energy data involves a wide range of data and a large amount of data, the amount of data associated with an energy data sequence may be difficult to estimate. If no corresponding processing is performed, a large amount of computing resources will be consumed. Therefore, the remaining data with weak correlation with the feature fusion target data sequence should be eliminated to ensure the efficiency and accuracy of feature fusion.

[0053] Specifically, taking the b-th energy data sequence of the a-th energy data as an example, the b-th energy data sequence is used as the input of the polynomial curve fitting algorithm, and the output is the polynomial curve of the b-th energy data sequence, and the peak points therein are obtained. The polynomial curve fitting algorithm is a well-known technology, and the process is not repeated in this embodiment. The mean of all the peak points in the curve is calculated, and the peak points on the polynomial curve with values ​​greater than or equal to the mean of the peak points are taken as super-standard peak points, and the peak points with values ​​less than the mean of the peak points are taken as low-standard peak points. Under normal circumstances, the appearance of super-standard peak points of energy data is usually the peak period of energy consumption, with a low frequency of occurrence and a short duration; while the low-standard peak is a small fluctuation of conventional energy consumption data, with a high frequency of occurrence, and usually continuous, with a relatively small fluctuation amplitude.

[0054] All the super-standard peak points in the polynomial curve of the b-th energy data sequence are used as the input of the regional growing algorithm, and each super-standard peak point is used as the initial growing point. The growth criterion of the regional growing algorithm is that when the number of sampling moments between the initial point and the target point is 1, growth is performed. The stopping condition is that when there is no data point that meets the growth criterion in the neighborhood of the growing point, growth is stopped. The output is multiple growing regions. The super-standard peak points in each growing region are sorted in ascending order according to the sampling time to construct a super-standard peak point data sequence; all the low-standard peak points in the polynomial curve of the b-th energy data sequence are used as the input of the regional growing algorithm, and each low-standard peak point is used as the initial growing point. The growth criterion of the regional growing algorithm is that when the number of sampling moments between the initial point and the target point is 1, growth is performed. The stopping condition is that when there is no data point that meets the growth criterion in the neighborhood of the growing point, growth is stopped. The output is multiple growing regions. The low-standard peak points in each growing region are sorted in ascending order according to the sampling time sequence to construct a low-standard peak point data sequence. The regional growing algorithm is a well-known technology, and its process is not repeated in this embodiment. Construct the energy peak connection intensity index of the b-th energy data series:

[0055]

[0056]

[0057]

[0058] Where D b is the significance index of exceeding the standard for the bth energy data series, E is the total number of data series of exceeding the standard peak points in the bth energy data series, and F is the total number of data series of substandard peak points in the bth energy data series. is the mean of all data in the jth super-standard peak point data sequence of the bth energy data sequence, is the mean of all data in the i-th low standard peak data sequence of the b-th energy data sequence, len(*) is the function for obtaining the length of the data sequence, is the jth super-standard peak point data sequence of the bth energy data sequence, is the length of the jth super-standard peak point data sequence of the bth energy data sequence, is the length mean of all low-standard peak point data sequences and all super-standard peak point data sequences of the b-th energy data sequence;

[0059] C b is the characteristic disturbance index of the b-th energy data sequence, is the mean of all data in the kth low standard peak data sequence of the bth energy data sequence, are the i-th and k-th low standard peak data sequences of the b-th energy data sequence, are the lengths of the i-th and k-th low standard peak data sequences of the b-th energy data sequence respectively;

[0060] H b is the energy peak connection intensity index of the bth energy data sequence, Z b , Z g are the bth and gth energy data sequences respectively, cos(*) is the cosine similarity function, cos(Z b ,Z g ) is to obtain the cosine similarity between the bth and gth energy data sequences, h is the total number of energy data sequences of the ath type of energy data, C g is the characteristic disturbance index of the g-th energy data sequence, and α is a parameter to prevent the denominator from being 0. It should be noted that the value of α can be set by the implementer. In this embodiment, the value of α is set to 0.001. is the first sum value.

[0061] When the peak difference between the super-standard and sub-standard peak point data sequences in the b-th energy data is larger, and the average length of the super-standard peak point data sequence is longer, The larger the value is, the more significant the exceeding of the standard is in the energy data series, that is, D b When the mean difference and length difference of the low standard peak data sequence in the bth energy data sequence are smaller, the number of occurrences is greater, and the significance of exceeding the standard is stronger, C b The larger the value of , the easier it is for the standard characteristic conditions in the energy data sequence to be disturbed, and the more likely it is to exceed the standard. When the cosine similarity between the bth energy data sequence and the other energy data sequences is larger, and the difference in the characteristic disturbance index is smaller, it means that the correlation between the energy data sequence and the other energy data sequences is greater, that is, H b The bigger.

[0062] At this point, the energy peak connection intensity index of each energy data series is obtained.

[0063] The energy peak connection strength index of all energy data sequences of the same type of energy data is used as the input of the K-means clustering algorithm, the number of clusters is set to 2, and the output is two clusters. The mean of the energy peak connection strength index in each cluster is calculated. Due to the energy peak connection strength, the cluster with the smallest mean of the energy peak connection strength index is regarded as a weak connection cluster, and all energy data sequences in the cluster are eliminated; the cluster with the largest mean of the energy peak connection strength index is regarded as a strong connection cluster, and all energy data sequences in the cluster are regarded as strong connection energy data sequences. Each strongly connected energy data sequence is used as the input of the autoregressive moving average model (ARIMA forecasting algorithm), and the ARIMA forecasting algorithm is used to predict 50 data of all strongly connected energy data sequences. The output of the ARIMA forecasting algorithm is the forecast data from the n+1th moment to the n+51th moment of all strongly connected energy data sequences, where the nth moment is the moment corresponding to the last data of the strongly connected energy data sequence, and the forecast data from the n+1th moment to the n+51th moment are arranged in ascending order according to the time sequence, and the forecast feature sequence of each strongly connected energy data sequence is constructed and recorded as in The meaning is the prediction feature sequence of the b′th strongly connected energy data sequence. The K-means clustering algorithm and the ARIMA prediction algorithm are both well-known technologies, and the process will not be described in detail in this embodiment.

[0064] Data prediction algorithms are usually based on the changing patterns of existing data and then make predictions based on certain influencing factors. That is, the predicted data is strongly correlated with the changes in the existing data, so the development of the predicted value of the data sequence is closely related to the data characteristics of the data sequence itself. The stronger the correlation between the two data sequences, the stronger the correlation between their predicted feature sequences. Each predicted feature sequence is used as the input of the ApEn algorithm, and the output is the ApEn value of each predicted feature sequence. The ApEn algorithm is a well-known technology, and its process is not repeated in this embodiment. All strongly connected energy data sequences are then used as inputs for the autocorrelation coefficients, and the interval parameter of the autocorrelation coefficient is set to half the length of the strongly connected energy data sequence (for example, if the length of the strongly connected energy data sequence is n, then the autocorrelation coefficient interval parameter is ), the output is the autocorrelation coefficient of each strongly connected energy data series.

[0065] For the prediction data of each strongly connected energy data sequence, starting from the last data of each strongly connected energy data sequence, each prediction result is added to the strongly connected energy data sequence one by one, and the energy data sequence at each prediction moment is constructed and recorded as in The meaning is the energy data sequence at the yth prediction moment of the b′th strongly connected energy data sequence. For example, if the total number of current strongly connected energy data sequences is n, and there are m current prediction data, then the length of the energy data sequence at the current prediction moment is n+m. If there are a current prediction data, then the length of the energy data sequence at the current prediction moment is n+a, and so on.

[0066] Then, the energy data sequence at each prediction moment is used as the input of the autocorrelation coefficient, and the interval parameter of the autocorrelation coefficient is set to half the length of the energy data sequence at the prediction moment (if the length of the energy data sequence at the prediction moment is n, then the autocorrelation coefficient interval parameter is ), the output is the autocorrelation coefficient of the energy data sequence at each prediction moment. The autocorrelation coefficient is a well-known technology, and this embodiment will not repeat its process. Constructing the prediction feature sequence The prediction feature correlation coefficient of is:

[0067]

[0068]

[0069]

[0070] In the formula, To predict the feature sequence The prediction reliability weight of the mth prediction data, T(*) is the acquisition time corresponding to the data point, To predict the feature sequence The mth predicted data in To predict the feature sequence The mth predicted data corresponds to the collection time, N m To predict the feature sequence The data value of the mth predicted data in , is the mean of all data in the b′th strongly connected energy data sequence;

[0071] To predict the feature sequence The prediction excellence index of the mth prediction data, N l To predict the feature sequence The data value of the lth predicted data in, X is the predicted feature sequence The number of predicted data in , ρ nis the autocorrelation coefficient of the b′th strongly connected energy data sequence, ρ n+m is the autocorrelation coefficient of the energy data sequence at the mth prediction moment of the strongly connected energy data sequence, γ is the parameter to prevent the denominator from being 0, recorded as the first parameter;

[0072] L b′ To predict the feature sequence The prediction feature correlation coefficient, u is the total number of strongly connected energy data sequences, ApEn b′ ,ApEn g′ To predict the feature sequence and the predicted feature sequence ApEn value, e * is an exponential function with the natural constant e as base, To predict the feature sequence The prediction excellence index of the mth prediction data in , To prevent the parameter with a denominator of 0, it is recorded as the second parameter. It should be noted that γ and The implementer can set the value of , and in this embodiment, γ and The values ​​of are all set to 0.001.

[0073] when sequence The smaller the time of the mth forecast data is, and the smaller the difference between the forecast data and the data mean of the existing energy data series is, The larger the value of is, the greater the correlation between the predicted value and the existing energy data series is, and the higher the credibility is. The greater the difference between the mth predicted data and the rest of the data in the predicted data sequence, and the smaller the difference between the autocorrelation coefficient of the energy data sequence before the prediction and the autocorrelation coefficient of the energy data sequence when predicting to the mth data, The larger the value, the better the prediction of the forecast data is and the more it fits the characteristics of the existing energy data. The closer the ApEn value of L is to the ApEn values ​​of the other prediction feature sequences, and the more similar the prediction excellence of each data in the prediction feature sequence is, the better the prediction performance of L is. b′ The larger the value of , the higher the correlation between the prediction characteristics of the energy data series and other energy data series.

[0074] At this point, the predicted feature correlation coefficient of each predicted feature sequence is obtained.

[0075] Step S003, dividing the energy data sequences that are truly strongly correlated according to the predicted feature correlation coefficient, and then using deep learning to fuse the context information features.

[0076] The prediction feature correlation coefficients of all prediction feature sequences are used as the input of the K-mediods clustering algorithm, the number of clusters is set to 2, and the output is two clusters. The mean of the prediction feature correlation coefficients in each cluster is calculated, and the cluster with the smallest mean of the prediction feature correlation coefficient is taken as the non-correlated cluster, and all the strongly connected energy data sequences are eliminated; the cluster with the largest mean of the prediction feature correlation coefficient is taken as the related cluster.

[0077] Since the prediction feature correlation coefficient of each prediction feature sequence in the associated clustering cluster is large, it means that each energy data sequence in the associated clustering cluster has a strong correlation with other energy data sequences. The energy data sequence corresponding to each prediction feature sequence in the associated clustering cluster is recorded as a high-correlation energy data sequence. Taking the b″th highly correlated energy data sequence as an example, the b″th highly correlated energy data sequence and other highly correlated energy data sequences are used as the input of the autoencoder, the activation function of the autoencoder is the sigmoid function, the loss function is the mean square error, and the output is the sequence features of the b″th highly correlated energy data sequence and other highly correlated energy data sequences. Then, the sequence features of the b″th highly correlated energy data sequence and the sequence features of other highly correlated energy data sequences are used as input, and the reinforcement learning strategy feature fusion method is adopted, and the output is the feature fusion result. Finally, the context feature fusion result of the b″th highly correlated energy data sequence is obtained. Among them, the reinforcement learning strategy feature fusion method is a well-known technology, and its process is not repeated in this embodiment.

[0078] So far, a method for fusion of energy data context information features based on reinforcement learning has been completed. The steps of the above method are shown in the figure below. Figure 2 shown.

[0079] In summary, the embodiment of the present invention obtains the feature disturbance index and the exceeding significance index by analyzing the prominence of the energy data peak exceeding the standard and the susceptibility of the data to interference, constructs the energy peak connection strength index, and preliminarily eliminates a part of the data with weak correlation; predicts the energy data sequence through the ARIMA prediction algorithm to obtain the prediction feature sequence; constructs the prediction feature correlation coefficient by analyzing the prediction effect of the data in the prediction feature sequence and the degree of correlation with the original data, further determines the data with strong correlation with the energy data sequence currently required to be analyzed, reduces the amount of data when performing context information feature fusion, and then divides the data sequence that really has strong correlation, and then uses deep learning to perform context information feature fusion; solves the problem that the traditional feature fusion algorithm only considers the correlation but not the essence of the correlation, resulting in too much correlated data, and ultimately leads to low accuracy of the feature after feature fusion, thereby improving the accuracy of energy data context feature information fusion.

[0080] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0081] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0082] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Modifications to the technical solutions recorded in the aforementioned embodiments, or equivalent replacement of some of the technical features therein, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for fusion of energy data context information features based on reinforcement learning, characterized in that: The method comprises the following steps: Collect each energy data sequence in each type of energy data; The polynomial curve of each energy data sequence is obtained by a polynomial curve fitting algorithm, and the super-standard and sub-standard peak point data sequences of each energy data sequence are obtained according to the peak point size and corresponding time in the polynomial curve; in the same type of energy data, the energy peak connection intensity index of each energy data sequence is obtained according to the data changes in the super-standard and sub-standard peak point data sequences; the energy peak connection intensity index of all energy data sequences is clustered by a clustering algorithm to obtain a strongly connected energy data sequence, and the prediction feature sequence of each strongly connected energy data sequence is obtained by combining the ARIMA prediction algorithm; the prediction reliability weight of each data in the prediction feature sequence is obtained according to the data changes in the strongly connected energy data sequence and the prediction feature sequence; the prediction excellence index of each data in the prediction feature sequence is constructed according to the prediction reliability weight and the autocorrelation changes before and after the data in the prediction feature sequence is added to the strongly connected energy data sequence; the prediction feature correlation coefficient of each prediction feature sequence is obtained according to the prediction excellence index and the data difference in different prediction feature sequences; According to the prediction feature correlation coefficient combined with the clustering algorithm, each highly correlated energy data sequence is obtained, and the context feature fusion result of each highly correlated energy data sequence is obtained by the reinforcement learning strategy feature fusion method; The energy peak connection intensity index of each energy data sequence is obtained according to the data changes in the super and substandard peak point data sequence, specifically: The characteristic disturbance index of each energy data series is obtained according to the numerical difference and length change in each over-standard and under-standard peak data series; Calculate the cosine similarity between the b-th energy data sequence and each energy data sequence; calculate the sum of all the cosine similarities of the b-th energy data sequence; calculate the absolute value of the difference between the characteristic susceptibility index of the b-th energy data sequence and each energy data sequence; calculate the sum of all the absolute values ​​of the difference of the b-th energy data sequence; calculate the sum of the absolute values ​​of the difference and the sum of the preset parameters, recorded as the first sum; take the ratio of the sum of the cosine similarities to the first sum as the energy peak connection intensity index of the b-th energy data sequence; The prediction reliability weight of each data in the prediction feature sequence is obtained according to the data changes in the strongly connected energy data sequence and the prediction feature sequence, specifically including: Calculate the mean of all data in each strongly connected energy data sequence; calculate the absolute value of the difference between each predicted data and the mean in the predicted feature sequence of each strongly connected energy data sequence; calculate the calculation result of an exponential function with a natural constant as the base and the opposite number of the absolute value of the difference as the exponent; and use the ratio of the calculation result to the corresponding collection time of each predicted data as the prediction reliability weight of each predicted data; The construction of the prediction excellence index of each data in the prediction feature sequence specifically includes: Calculate the autocorrelation coefficient ρ of the b′th strongly connected energy data sequence n ; Calculate the autocorrelation coefficient ρ of the energy data sequence at the mth prediction moment of the b′th strongly connected energy data sequence n+m ; The specific expression for calculating the prediction excellence index of each data in the prediction feature sequence is: In the formula, is the prediction excellence index of the mth prediction data in the b′th prediction feature sequence, is the prediction reliability weight of the mth prediction data in the b′th prediction feature sequence, X is the number of prediction data in the b′th prediction feature sequence, N m and N l are the data values ​​of the mth and lth prediction data in the b′th prediction feature sequence, respectively, and γ is the preset first parameter; The prediction feature correlation coefficient of each prediction feature sequence is obtained according to the prediction excellence index and the data difference in different prediction feature sequences, specifically including: The ApEn value of each prediction feature sequence is obtained by the ApEn algorithm; the specific expression for calculating the prediction feature association coefficient of each prediction feature sequence is: Where, L b′ is the prediction feature correlation coefficient of the b′th prediction feature sequence, e * is an exponential function with the natural constant e as the base, u is the total number of strongly connected energy data sequences, X is the number of predicted data in the b′th predicted feature sequence, are the prediction excellence index of the mth prediction data in the b′th and g′th prediction feature sequences, respectively. θ is the preset second parameter. ApEn b′ ,ApEn g′ are the ApEn values ​​of the b′th and g′th predicted feature sequences.

2. The energy data context information feature fusion method based on reinforcement learning as claimed in claim 1 is characterized in that: The polynomial curve of each energy data sequence is obtained by a polynomial curve fitting algorithm, and the super and substandard peak point data sequences of each energy data sequence are obtained according to the peak point size and corresponding time in the polynomial curve, specifically: For each energy data sequence, a polynomial curve of the energy data sequence is obtained by fitting the energy data sequence through a polynomial curve fitting algorithm; the mean of all peak point data in the polynomial curve is calculated; the peak point with a value greater than the mean is regarded as an over-standard peak point, and the peak point with a value less than the mean is regarded as a low-standard peak point; All the super-standard peak points on the polynomial curve are used as the input of the regional growing algorithm, where each super-standard peak point is used as the initial growth point, and the algorithm outputs each growth area; the sequence composed of all the super-standard peak points in each growth area is used as the super-standard peak point data sequence of the energy data sequence; and each low-standard peak point data sequence is obtained by obtaining each super-standard peak point data sequence.

3. The energy data context information feature fusion method based on reinforcement learning as claimed in claim 1, characterized in that: The characteristic disturbance susceptibility index of each energy data sequence is obtained according to the difference and length change between each super-standard and sub-standard peak point data sequence, and the expression is: In the formula, C b is the characteristic disturbance index of the b-th energy data sequence, D b is the significance index of exceeding the standard for the bth energy data series, F and E are the total number of low-standard peak data series and exceeding-standard peak data series of the bth energy data series, respectively. are the means of all data in the i-th and k-th low standard peak data sequences of the b-th energy data sequence, are the ith and kth low standard peak data sequences of the bth energy data sequence, respectively. len(*) is the function for obtaining the length of the data sequence. are the lengths of the i-th and k-th low standard peak data sequences of the b-th energy data sequence, is the mean of all data in the jth super-standard peak point data sequence of the bth energy data sequence, is the jth super-standard peak point data sequence of the bth energy data sequence, is the length mean of all low-standard peak point data sequences and all super-standard peak point data sequences of the b-th energy data sequence; is the length of the j-th super-standard peak point data sequence of the b-th energy data sequence.

4. The energy data context information feature fusion method based on reinforcement learning according to claim 1 is characterized in that: The energy peak connection intensity index of all energy data sequences is clustered by a clustering algorithm to obtain a strongly connected energy data sequence, and the prediction feature sequence of each strongly connected energy data sequence is obtained by combining the ARIMA prediction algorithm, specifically: The energy peak connection intensity index of all energy data sequences is clustered by clustering algorithm to obtain two clusters; the mean of all elements in each cluster is calculated; the cluster with the largest mean is taken as the strong connection cluster; The energy data sequences in the strong connection clusters are regarded as the strong connection energy data sequences; The ARIMA prediction algorithm is used to predict the data of each strongly connected energy data series for a preset number of moments, and the sequence composed of all the predicted data of each strongly connected energy data series is used as the prediction feature sequence of each strongly connected energy data series.

5. The energy data context information feature fusion method based on reinforcement learning according to claim 1 is characterized in that: The energy data sequence at the prediction time is specifically: The 1st to yth predicted data of the strongly connected energy data sequence are added to the strongly connected energy data sequence, and the formed new sequence is used as the yth predicted moment energy data sequence of the strongly connected energy data sequence.

6. The energy data context information feature fusion method based on reinforcement learning according to claim 1 is characterized in that: The prediction feature correlation coefficient is combined with the clustering algorithm to obtain each highly correlated energy data sequence, and the context feature fusion result of each highly correlated energy data sequence is obtained by the reinforcement learning strategy feature fusion method, specifically: Perform K-mediods clustering on the predicted feature correlation coefficients of all predicted feature sequences to obtain two clusters, and take the cluster with the largest mean of all predicted feature correlation coefficients in the cluster as the correlation cluster; The energy data sequences corresponding to all prediction feature sequences in the associated clusters are recorded as high-correlation energy data sequences; each high-correlation energy data sequence is used as the input of the autoencoder, and the output is the sequence features of each high-correlation energy data sequence; The sequence features of the b″th highly correlated energy data sequence and the sequence features of other highly correlated energy data sequences are taken as input, and the feature fusion method of reinforcement learning strategy is adopted, and the output is the feature fusion result.

Images