A Deep Learning Model Analysis Method and System for Predicting New Energy Power Generation Output

Abstract

This invention discloses a method and system for analyzing deep learning models for predicting renewable energy power generation output. The method includes: constructing and training a deep learning-based renewable energy power generation output prediction model; simplifying the input features of the trained prediction model to extract key influencing factors; calculating the contribution of these key influencing factors to analyze their impact on the prediction results; classifying the power generation data samples by meteorological type based on the clear sky index; transforming the regression layer of the prediction model into a logical inference layer; fine-tuning the parameters of the logical inference layer of the prediction model based on the meteorological type-classified power generation data samples, and analyzing the output logic of the prediction model. This invention improves the reliability of deep learning models for predicting renewable energy power generation output, assists power operators in understanding the predictive inference logic of deep learning models, and enhances the safety and stability of the power system in absorbing renewable energy power generation.

Description

Technical Field

[0001] This invention relates to a deep learning model analysis method and system for predicting the output of new energy power generation, belonging to the field of power system big data analysis technology. Background Technology

[0002] The penetration rate of renewable energy generation in new power systems is increasing year by year, which makes the randomness and volatility of renewable energy generation a major factor affecting the stability of new power systems. Therefore, developing high-precision renewable energy output forecasting technology is of great significance for improving the safe and stable operation of power systems and promoting the green and low-carbon transformation of energy.

[0003] For renewable energy power generation forecasting technology, current mainstream methods mainly employ models categorized into simple statistical models, traditional machine learning models, and deep learning models. Among these, deep learning models, based on deep neural networks, are better able to fit high-dimensional nonlinear problems, thus demonstrating superior performance in various multi-timescale renewable energy power generation forecasts. However, the improved prediction accuracy of deep learning models comes at the cost of becoming deeper and more complex, resulting in extremely poor interpretability. Especially when extreme prediction error events occur, the poor interpretability of deep learning models makes it difficult to provide effective guidance for adjusting renewable energy consumption strategies. Therefore, there is an urgent need to develop analytical methods for deep learning models used in renewable energy power generation forecasting. Summary of the Invention

[0004] Purpose of the invention: This invention addresses the shortcomings of current power system big data analysis technology in new energy power generation forecasting methods. It proposes a deep learning model analysis method and system for predicting new energy power generation output. By analyzing the key influencing factors and output logic of the deep learning model, the performance of the deep learning model for predicting new energy power generation output is improved. This helps guide power system dispatchers to adjust new energy power generation consumption strategies in a timely manner, thereby increasing the power system's new energy consumption capacity and enhancing the competitiveness of power grid companies.

[0005] Technical Solution: To achieve the above-mentioned objectives, the present invention proposes a deep learning model analysis method for predicting the output of new energy power generation, comprising the following steps:

[0006] Step 1: Construct a deep learning-based prediction model for renewable energy power generation output. Train the prediction model based on renewable energy power generation output data samples. The input features of the prediction model and data samples include historical time series of renewable energy power generation, historical meteorological time series, weather forecasts, and timestamps. The formula is expressed as:

[0007]

[0008] Where, The predicted output value of new energy power generation from time t+1 to t+L; f pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions;

[0009] Step 2: Simplify the input features of the trained prediction model and extract the key influencing factors. Among them, 1+W+S represents the number of key influencing factors. The range of the real number field of dimension 1+W+S;

[0010] Step 3: Calculate the contribution value of key influencing factors using the Shapley summation interpretation algorithm, and analyze the degree of influence of key influencing factors on the prediction results. The analytical formula is expressed as:

[0011]

[0012] x j ∈x,1≤i≤L,1≤j≤(1+W+S)

[0013] Where, Here, φ0 represents the predicted output of new energy power generation at time t+i, and x represents the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value;

[0014] Step 4: Classify the meteorological types of the new energy power generation output data samples based on the clear sky index;

[0015] Step 5: Transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, fine-tune the parameters of the logical inference layer of the prediction model, and parse out the output logic of the prediction model. The analytical formula is expressed as:

[0016]

[0017] In the formula, C t+iLet p(C) be the weather type of the data sample at time t+i. t+i ==k) is the probability that the weather type of the data sample at time t+i belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class for the data sample at time t+i.

[0018] Preferably, step 2 involves simplifying the input features of the trained prediction model to extract key influencing factors. The specific implementation steps are as follows:

[0019] An approximation function for the input features is constructed for the trained prediction model, expressed by the formula:

[0020]

[0021] Where x wp,t+i For a weather forecast at time t+i, the domain is a real number field of dimension 1×W. x ts,t+i Let be the timestamp at time t+i, contained in a real number field of dimension 1×S. Φ1 is an approximate function of the historical time series of new energy power generation, which is contained in the real number domain. Φ2 is an approximate function for meteorology, contained within the real number domain of dimension W. Φ3 is an approximate function of the timestamp, contained within the real number domain of dimension S.

[0022] Based on the objective of minimizing prediction and fitting errors, an approximation function of the input features is solved. The solved approximation functions are then combined as column vectors to obtain the simplified key influencing factors, expressed by the following formula:

[0023] x = f concat (Φ1,Φ2,Φ3);

[0024] In the formula, f concat This is a column vector merging function, where x is the key influencing factor.

[0025] As a preferred option, the objective function formula for solving the input feature approximation function is:

[0026]

[0027] In the formula, P t+i This represents the actual output value of new energy power generation at time (t+i). It is the inverse function of Φ1. It is the inverse function of Φ2. It is the inverse function of φ3. This is the sum of squares of the errors.

[0028] Preferably, step 3 uses the Shapley summation interpretation algorithm to calculate the contribution value of the key influencing factors. The calculation formula for this algorithm is as follows:

[0029]

[0030] Where x j For the j-th key impact factor, g i Key Influencing Factors The contribution value, x\{x j} does not contain x j The set of key influencing factors, S is a subset of x\{x j A subset of}, S∪{x} j} is a subset S containing x j The set after, || is the length of the set, ! is the factorial, x S It is a key influencing factor in S. For S∪{x j Key influencing factors in} For input x S Deep learning models, For input The deep learning model.

[0031] Preferably, the meteorological type classification of the new energy power generation output data sample based on the clear sky index in step 4 is expressed by the following formula:

[0032]

[0033] In the formula, C t Let k be the weather type of the data sample at time t, k be the weather type number, and K be the total number of weather types. If (C t ==k) is the conditional judgment that the weather type of the data sample at time t belongs to the kth category, P t a P represents the actual output of new energy power generation for the data sample at time t. t clear The data sample represents the clear-sky value of renewable energy power generation output at time t. The clear sky index.

[0034] Preferably, step 5 involves transforming the regression layer of the new energy power generation output prediction model into a logical reasoning layer. The specific implementation steps are as follows:

[0035] The number of output nodes in the regression layer of the new energy power generation output prediction model is set to the total number of weather types, K. A K-category classification function is constructed, expressed by the formula: p(C t+i ==k)=fclass (H D ×W D×K [k];1≤k≤K,where C t+i Let p(C) be the weather type of the data sample at time (t+i). t+i =k) ​​is the probability that the weather type of the data sample at time (t+i) belongs to the kth class, f class H is the classification function. D Here, D represents the number of input nodes in the regression layer, and W represents the number of input nodes. D×K Here, [k] represents the weights of the classification function, and [k] represents the output node number of the classification function.

[0036] The weights W of the classification function D×K The weight vector in the formula is set as the end node of the logic inference layer, and is expressed as: n i =w k ∈W D×K ; 1≤i≤K, where n i w is the i-th terminal node of the logic reasoning layer. k For W D×K The k-th weight vector in;

[0037] Calculate the inference nodes one level above the terminal node, up to ∑ j K(j) = 1, the last n j This is the first node of the logical reasoning layer;

[0038] Copy the K-class classification function, randomly initialize the weights of the classification function, and construct the classification weight function, expressed by the formula: 1≤k≤K, where h k,t+i Let be the probability weight of a data sample belonging to the k-th category of weather type at time t+i. H is the classification weight function. D As the input node of the regression layer, [k] represents the weights of the classification function, which are randomly initialized. [k] represents the output node number of the classification function.

[0039] Preferably, the inference node of the layer above the terminal node is calculated, and the formula is expressed as:

[0040]

[0041] In the formula, n j K is the j-th inference node at the level above the end node, and K(j) is the number of end nodes under the j-th inference node.

[0042] This invention also provides a deep learning model analysis system for predicting the output of new energy power generation, comprising:

[0043] The model building and training module is used to construct a deep learning-based prediction model for new energy power generation output. The model is trained based on new energy power generation output data samples. The input features of the prediction model and data samples include historical time series data of new energy power generation, historical meteorological time series data, weather forecasts, and timestamps. The formula is expressed as follows:

[0044]

[0045] Where, The predicted output value of new energy power generation from time t+1 to t+L; f pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions;

[0046] The input feature simplification module is used to simplify the input features of the trained prediction model and extract key influencing factors. Among them, 1+W+S represents the number of key influencing factors. The range of the real number field of dimension 1+W+S;

[0047] The impact analysis module uses the Shapley summation interpretation algorithm to calculate the contribution value of key impact factors, and analyzes the degree of influence of key impact factors on the prediction results. The analytical formula is expressed as follows:

[0048]

[0049] x j ∈x,1≤i≤L,1≤j≤(1+W+S)

[0050] Where, Here, φ0 represents the predicted output of new energy power generation at time t+i, and x represents the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value;

[0051] The meteorological type classification module is used to classify the meteorological type of new energy power generation output data samples based on the clear sky index.

[0052] The model output parsing module is used to transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, the parameters of the logical inference layer of the prediction model are fine-tuned, and the output logic of the prediction model is parsed out. The parsing formula is expressed as follows:

[0053]

[0054] In the formula, C t+i Let p(C) be the weather type of the data sample at time t+i. t+i ==k) is the probability that the weather type of the data sample at time t+i belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class at time t+i.

[0055] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the deep learning model parsing method for predicting the output of new energy power generation as described above.

[0056] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning model analysis method for predicting the output of new energy power generation as described above.

[0057] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0058] (1) This invention constructs a deep learning-based new energy power generation output prediction model and trains the model; it then simplifies the input features of the trained model to extract key influencing factors; it calculates the contribution of these key influencing factors to analyze their impact on the prediction results; it classifies the power generation data samples by meteorological type based on the clear sky index; it transforms the regression layer of the prediction model into a logical inference layer; and it fine-tunes the parameters of the logical inference layer of the prediction model based on the output data samples classified by meteorological type, thus analyzing the output logic of the prediction model. Compared with existing model analysis methods, the method of this invention can better adapt to deep learning models and has higher accuracy in predicting new energy power generation output.

[0059] (2) The method of the present invention can meet the forecasting needs of various new energy power generation output based on numerical weather forecasting, and improve the interpretability and reliability of the new energy power generation output forecasting model.

[0060] (3) The model analysis results of the method of the present invention can provide guidance and data support for dispatchers of new power systems, thereby improving the power system's ability to absorb new energy power generation and improving the economic efficiency of power system operation. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the principle of the method of the present invention.

[0062] Figure 2 This is a graph showing the analysis results of key influencing factors of the method of this invention.

[0063] Figure 3 This is a diagram showing the output logic analysis results of the method of the present invention. Detailed Implementation

[0064] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0065] like Figure 1 As shown, this invention proposes a deep learning model analytical method for predicting the output of new energy power generation, which specifically includes the following steps:

[0066] Step 1: Construct a deep learning-based prediction model for renewable energy power generation output. Train the prediction model based on renewable energy power generation output data samples. The input features of the prediction model and data samples include historical time series of renewable energy power generation, historical meteorological time series, weather forecasts, and timestamps. The formula is expressed as:

[0067]

[0068] Where, f represents the predicted output of new energy power generation from time (t+1) to (t+L). pred For deep learning models, x pv The historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within a real number domain of dimension D×W. x wp For weather forecasts, the real number domain encompassed by dimensions L×W is defined. x ts A timestamp, contained in a real number field of dimension L×S. L is the length of the predicted output value, D is the length of the historical time series, W is the number of meteorological attributes, and S is the number of timestamp dimensions.

[0069] Step 2: Simplify the input features of the trained prediction model and extract the key influencing factors. Where (1+W+S) represents the number of key influencing factors, The range of the real number field of dimension (1+W+S)

[0070] Step 3: Calculate the contribution value of key influencing factors using the Shapley summation interpretation algorithm, and analyze the degree of influence of key influencing factors on the prediction results. The analytical formula is expressed as:

[0071]

[0072] x j ∈x,1≤i≤L,1≤j≤(1+W+S)

[0073] Where, Let φ0 be the predicted output of new energy power generation at time (t+i), and x be the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value;

[0074] Step 4: Classify the meteorological type of the renewable energy power generation output data sample based on the clear sky index. The formula is as follows:

[0075]

[0076] In the formula, C t Let k be the weather type of the data sample at time t, k be the weather type number, and K be the total number of weather types. If (C t ==k) is the conditional judgment that the weather type of the data sample at time t belongs to the kth category, P t a P represents the actual output of new energy power generation for the data sample at time t. t clear The data sample represents the clear-sky value of renewable energy power generation output at time t. The clear sky index;

[0077] Step 5: Transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, fine-tune the parameters of the logical inference layer of the prediction model, and parse out the output logic of the prediction model. The analytical formula is expressed as:

[0078]

[0079] In the formula, C t+i Let p(C) be the weather type of the data sample at time (t+i). t+i ==k) is the probability that the weather type of the data sample at time (t+i) belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class at time (t+i) of the data sample.

[0080] The following detailed implementation process of using the method of this invention to analyze a deep learning model for predicting the output of new energy power generation is described in detail with reference to specific embodiments. This invention selects distributed photovoltaic data from a county in Jiangsu Province in 2018 for predicting the output of new energy power generation over the past 24 hours. Therefore, the length of the predicted output value is selected as 24, the length of the historical time series is selected as 168, the number of meteorological attributes is selected as 8 (temperature, humidity, wind speed, wind direction, cloud cover, total radiation, direct radiation, diffuse radiation), and the number of timestamp dimensions is selected as 2 (hourly, daily). Based on the above data, the specific implementation steps of the method of this invention are as follows:

[0081] First, a deep learning-based prediction model for renewable energy power generation output is constructed. This model is trained using renewable energy power generation output data samples. The input features of the prediction model and the data samples include historical time series data of renewable energy power generation, historical meteorological time series data, weather forecasts, and timestamps. The formula is expressed as follows:

[0082]

[0083] In the formula, f represents the predicted output of new energy power generation from time (t+1) to (t+L); pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions;

[0084] Secondly, the input features of the trained prediction model are simplified, and an approximate function of the input features is constructed, expressed by the formula:

[0085]

[0086] Where x wp,t+i For a weather forecast at time (t+i), the domain is a real number field of dimension 1×W. x ts,t+i Let (t+i) be the timestamp at time (t+i), contained within a real number field of dimension 1×S. Φ1 is an approximate function of the historical time series of new energy power generation, which is contained in the real number domain. Φ2 is an approximate function for meteorology, contained within the real number domain of dimension W. Φ3 is an approximate function of the timestamp, contained within the real number domain of dimension S. Based on the objective of minimizing both prediction and fitting errors, the approximate function of the input features is solved, and the objective function formula is:

[0087]

[0088] In the formula, P t+i This represents the actual output value of new energy power generation at time (t+i). It is the inverse function of Φ1. It is the inverse function of Φ2. It is the inverse function of Φ3. The sum of squares of the errors is used; the approximate input feature functions after solving are combined as column vectors to obtain the simplified key influence factors, expressed by the formula:

[0089] x = f concat (Φ1,Φ2,Φ3);

[0090] In the formula, f concat This is a column vector merging function, where x is the key influencing factor; key influencing factors are extracted in a simplified manner. Where (1+W+S) represents the number of key influencing factors, The range of the real number field of dimension (1+W+S);

[0091] Next, the contribution values ​​of key influence factors are calculated using the Shapley summation interpretation algorithm. The formula for this algorithm is as follows:

[0092]

[0093] Where x j For the j-th key impact factor, g i Key Influencing Factors The contribution value, x\{x j} does not contain x j The set of key influencing factors, S is a subset of x\{x j A subset of}, S∪{x} j} is a subset S containing x j The set after, || is the length of the set, ! is the factorial, x S It is a key influencing factor in S. For S∪{x j Key influencing factors in} For input x S Deep learning models, For input The deep learning model was used to analyze the influence of key influencing factors on the prediction results, and the analytical formula is expressed as follows:

[0094]

[0095] x j ∈x,1≤i≤L,1≤j≤(1+W+S)

[0096] In the formula, Let φ0 be the predicted output of new energy power generation at time (t+i), and x be the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value, such as Figure 2 This diagram shows the results of analyzing key influencing factors in a deep learning model for predicting new energy power generation output in an embodiment of the present invention. The deep learning model predicted a significant photovoltaic power ramp-up at 12:30 PM on that day. After analyzing the key influencing factors, the prediction was found to be primarily influenced by cloud cover in the weather forecast, thus demonstrating high reliability. Subsequent comparison with actual values ​​confirmed that the deep learning model's prediction at that moment was indeed quite accurate, consistent with the analysis results of the present invention. Therefore, the method of the present invention, by analyzing the degree of influence of key influencing factors on the prediction results, can improve the reliability of deep learning model predictions.

[0097] Then, based on the clear sky index, the meteorological type of the new energy power generation output data sample is classified, and the formula is expressed as:

[0098]

[0099] In the formula, C t Let k be the weather type of the data sample at time t, k be the weather type number, and K be the total number of weather types. In this embodiment of the invention, three weather types are selected: sunny, cloudy, and rainy, with a value of K = 3; if(C t ==k) is the conditional judgment that the weather type of the data sample at time t belongs to the kth category, P t a P represents the actual output of new energy power generation for the data sample at time t. t clear The data sample represents the clear-sky value of renewable energy power generation output at time t. The clear sky index.

[0100] Finally, the regression layer of the new energy power generation output prediction model is transformed into a logical reasoning layer. The number of output nodes of the regression layer of the new energy power generation output prediction model is set to the total number of weather types, K, and a K-category classification function is constructed, expressed by the formula:

[0101] p(C t+i ==k)=f class (H D ×W D×K [k];1≤k≤K

[0102] In the formula, C t+i Let p(C) be the weather type of the data sample at time (t+i). t+i =k) ​​is the probability that the weather type of the data sample at time (t+i) belongs to the kth class, f class H is the classification function. D Here, D represents the number of input nodes in the regression layer, and W represents the number of input nodes. D×K The weights W of the classification function are given by [k], where [k] is the output node number of the classification function. D×K The weight vector in the formula is set as the end node of the logic inference layer, and is expressed by the formula:

[0103] n i =w k ∈W D×K ; 1≤i≤K

[0104] In the formula, n i w is the i-th terminal node of the logic reasoning layer. k For W D×K The k-th weight vector in the matrix; calculate the inference node of the layer above the terminal node, expressed by the formula:

[0105]

[0106] In the formula, n j Let K(j) be the j-th inference node at the level above the terminal node, and K(j) be the number of terminal nodes under the j-th inference node; repeat the above steps until ∑ j K(j) = 1, the last n j This is the first node of the logical reasoning layer; the K-category classification function is copied, the weights of the classification function are randomly initialized, and a classification weight function is constructed, expressed by the formula:

[0107]

[0108] In the formula, h k,t+i Let be the probability weight of the weather type belonging to the k-th class at time (t+i). H is the classification weight function. D As the input node of the regression layer, Here, [k] represents the weights of the randomly initialized classification function, and [k] represents the output node number of the classification function. Based on the meteorological type-based data samples of new energy power generation output, the parameters of the logical inference layer of the prediction model are fine-tuned, and the output logic of the prediction model is parsed out. The analytical formula is expressed as:

[0109]

[0110] In the formula, C t+i Let p(C) be the weather type of the data sample at time (t+i). t+i ==k) is the probability that the weather type of the data sample at time (t+i) belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the k-th class at time (t+i), such as Figure 3 The diagram shows the output logic analysis results of a deep learning model for predicting renewable energy power generation. The deep learning model predicted a significant photovoltaic power ramp-up at 12:30 PM. After analysis using the output logic of this invention, the predicted value has an over 80% probability of being cloudy or rainy at that time, indicating high reliability. However, the probability of cloudy skies turning overcast or remaining cloudy afterward is less than 50%, indicating lower reliability. Post-event comparison with actual values ​​shows that the deep learning model's prediction at that moment was relatively accurate, while the prediction error increased significantly at subsequent times, consistent with the analysis results of this invention. Therefore, this invention's method, by analyzing the logic of the prediction output, can intuitively display the level of prediction reliability, thereby improving the interpretability of the deep learning model. Thus, the deep model analysis method of this invention can provide guidance and data support for dispatchers of new power systems in adjusting renewable energy power generation consumption strategies.

[0111] In summary, the deep learning model analysis method for predicting renewable energy power generation output of this invention can better adapt to deep learning models, has higher accuracy in predicting renewable energy power generation output, can meet the needs of various renewable energy power generation output prediction based on numerical weather forecasts, improves the interpretability and reliability of renewable energy power generation output prediction models, provides guidance and data support for dispatchers of new power systems, thereby improving the renewable energy power generation absorption capacity of the power system and improving the economic efficiency of power system operation.

[0112] Based on the same technical concept as the method embodiments, the present invention also provides a deep learning model analysis system for predicting the output of new energy power generation, comprising:

[0113] The model building and training module is used to construct a deep learning-based prediction model for new energy power generation output. The model is trained based on new energy power generation output data samples. The input features of the prediction model and data samples include historical time series data of new energy power generation, historical meteorological time series data, weather forecasts, and timestamps. The formula is expressed as follows:

[0114]

[0115] In the formula, The predicted output value of new energy power generation from time t+1 to t+L; f pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions;

[0116] The input feature simplification module is used to simplify the input features of the trained prediction model and extract key influencing factors. Among them, 1+W+S represents the number of key influencing factors. The range of the real number field of dimension 1+W+S;

[0117] The impact analysis module uses the Shapley summation interpretation algorithm to calculate the contribution value of key impact factors, and analyzes the degree of influence of key impact factors on the prediction results. The analytical formula is expressed as follows:

[0118]

[0119] x j ∈x,1≤i≤L,1≤j≤(1+W+S)

[0120] In the formula, Here, φ0 represents the predicted output of new energy power generation at time t+i, and x represents the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value;

[0121] The meteorological type classification module is used to classify the meteorological type of new energy power generation output data samples based on the clear sky index.

[0122] The model output parsing module is used to transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, the parameters of the logical inference layer of the prediction model are fine-tuned, and the output logic of the prediction model is parsed out. The parsing formula is expressed as follows:

[0123]

[0124] In the formula, C t+i Let p(C) be the weather type of the data sample at time t+i. t+i ==k) is the probability that the weather type of the data sample at time t+i belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class for the data sample at time t+i.

[0125] It should be understood that the deep learning model analysis system for predicting the output of new energy power generation in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0126] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the deep learning model parsing method for predicting the output of new energy power generation as described above.

[0127] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning model analysis method for predicting the output of new energy power generation as described above.

[0128] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A deep learning model analytical method for predicting the output of new energy power generation, characterized in that, Includes the following steps: Step 1: Construct a deep learning-based prediction model for renewable energy power generation output. Train the prediction model based on renewable energy power generation output data samples. The input features of the prediction model and data samples include historical time series of renewable energy power generation, historical meteorological time series, weather forecasts, and timestamps. The formula is expressed as: Where, The predicted output value of new energy power generation from time t+1 to t+L; f pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions; Step 2: Simplify the input features of the trained prediction model and extract the key influencing factors. Among them, 1+W+S represents the number of key influencing factors. The range of the real number field of dimension 1+W+S; Step 3: Calculate the contribution value of key influencing factors using the Shapley summation interpretation algorithm, and analyze the degree of influence of key influencing factors on the prediction results. The analytical formula is expressed as: x j ∈x,1≤i≤L,1≤j≤(1+W+S) Where, Here, φ0 represents the predicted output of new energy power generation at time t+i, and x represents the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value; Step 4: Classify the meteorological types of the new energy power generation output data samples based on the clear sky index; Step 5: Transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, fine-tune the parameters of the logical inference layer of the prediction model, and parse out the output logic of the prediction model. The analytical formula is expressed as: In the formula, C t+i Let p(C) be the weather type of the data sample at time t+i. t+i ==k) is the probability that the weather type of the data sample at time t+i belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class at time t+i.

2. The method according to claim 1, characterized in that, Step 2 involves simplifying the input features of the trained prediction model and extracting key influencing factors. The specific implementation steps are as follows: An approximation function for the input features is constructed for the trained prediction model, expressed by the formula: Where x wp,t+i For a weather forecast at time t+i, the domain is a real number field of dimension 1×W. x ts,t+i Let be the timestamp at time t+i, contained in a real number field of dimension 1×S. Φ1 is an approximate function of the historical time series of new energy power generation, which is contained in the real number domain. Φ2 is an approximate function for meteorology, contained within the real number domain of dimension W. Φ3 is an approximate function of the timestamp, contained within the real number domain of dimension S. Based on the objective of minimizing prediction and fitting errors, an approximation function of the input features is solved. The solved approximation functions are then combined as column vectors to obtain the simplified key influencing factors, expressed by the following formula: In the formula, f concat This is a column vector merging function, where x is the key influencing factor.

3. The method according to claim 2, characterized in that, The objective function formula for solving the input feature approximation function is: In the formula, P t+i This represents the actual output value of new energy power generation at time (t+i). It is the inverse function of Φ1. It is the inverse function of Φ2. It is the inverse function of Φ3. This is the sum of squares of the errors.

4. The method according to claim 1, characterized in that, Step 3 describes the calculation of the contribution value of key influence factors using the Shapley summation interpretation algorithm. The calculation formula for this algorithm is as follows: Where x j For the j-th key impact factor, g i Key Influencing Factors The contribution value, x\{x j } does not contain x j The set of key influencing factors, S is a subset of x\{x j A subset of}, S∪{x} j } is a subset S containing x j The set after, || is the length of the set, ! is the factorial, x S It is a key influencing factor in S. For S∪{x j Key influencing factors in} For input x S Deep learning models, For input The deep learning model.

5. The method according to claim 1, characterized in that, Step 4 involves classifying the meteorological type of the new energy power generation output data sample based on the clear sky index, as expressed by the following formula: In the formula, C t Let k be the weather type of the data sample at time t, k be the weather type number, and K be the total number of weather types. If (C t ==k) is the conditional judgment that the weather type of the data sample at time t belongs to the kth category, P t a P represents the actual output of new energy power generation for the data sample at time t. t clear The data sample represents the clear-sky value of renewable energy power generation output at time t. The clear sky index.

6. The method according to claim 1, characterized in that, Step 5 involves transforming the regression layer of the new energy power generation output prediction model into a logical reasoning layer. The specific implementation steps are as follows: The number of output nodes in the regression layer of the new energy power generation output prediction model is set to the total number of weather types, K. A K-category classification function is constructed, expressed by the formula: p(C t+i ==k)=f class (H D ×W D×K [k];1≤k≤K,where C t+i Let p(C) be the weather type of the data sample at time (t+i). t+i =k) ​​is the probability that the weather type of the data sample at time t+i belongs to the k-th class, f class H is the classification function. D Here, D represents the number of input nodes in the regression layer, and W represents the number of input nodes. D×K Here, [k] represents the weights of the classification function, and [k] represents the output node number of the classification function. The weights W of the classification function D×K The weight vector in the formula is set as the end node of the logic inference layer, and is expressed as: n i =w k ∈W D×K ; 1≤i≤K, where n i w is the i-th terminal node of the logic reasoning layer. k For W D×K The k-th weight vector in; Calculate the inference nodes one level above the terminal node, up to ∑ j K(j) = 1, the last n j This is the first node of the logical reasoning layer; Copy the K-class classification function, randomly initialize the weights of the classification function, and construct the classification weight function, expressed by the formula: 1≤k≤K, where h k,t+i Let be the probability weight of the weather type belonging to the k-th class at time (t+i). H is the classification weight function. D As the input node of the regression layer, [k] represents the weights of the classification function, which are randomly initialized. [k] represents the output node number of the classification function.

7. The method according to claim 6, characterized in that, The formula for calculating the inference node one level above the terminal node is as follows: In the formula, n j K is the j-th inference node at the level above the end node, and K(j) is the number of end nodes under the j-th inference node.

8. A deep learning model analysis system for predicting the output of new energy power generation, characterized in that, include: The model building and training module is used to construct a deep learning-based prediction model for new energy power generation output. The model is trained based on new energy power generation output data samples. The input features of the prediction model and data samples include historical time series data of new energy power generation, historical meteorological time series data, weather forecasts, and timestamps. The formula is expressed as follows: In the formula, The predicted output value of new energy power generation from time t+1 to t+L; f pred For deep learning models; x pv This is a historical time series of new energy power generation, contained within a real number field of dimension D×1. x w It is a historical meteorological time series, contained within the real number domain of dimension D×W. x wp For weather forecasting, it encompasses the real number domain of dimension L×W. x ts For timestamps, the data is contained within a real number field of dimension L×S. L is the length of the predicted output value; D is the length of the historical time series; W is the number of meteorological attributes; S is the number of timestamp dimensions; The input feature simplification module is used to simplify the input features of the trained prediction model and extract key influencing factors. Among them, 1+W+S represents the number of key influencing factors. The range of the real number field of dimension 1+W+S; The impact analysis module uses the Shapley summation interpretation algorithm to calculate the contribution value of key impact factors, and analyzes the degree of influence of key impact factors on the prediction results. The analytical formula is expressed as follows: Where, Here, φ0 represents the predicted output of new energy power generation at time t+i, and x represents the contribution bias value. j For the j-th key impact factor, g i Key Influencing Factors Contribution value; The meteorological type classification module is used to classify the meteorological type of new energy power generation output data samples based on the clear sky index. The model output parsing module is used to transform the regression layer of the new energy power generation output prediction model into a logical inference layer. Based on the new energy power generation output data samples classified by meteorological type, the parameters of the logical inference layer of the prediction model are fine-tuned, and the output logic of the prediction model is parsed out. The parsing formula is expressed as follows: In the formula, C t+i Let p(C) be the weather type of the data sample at time t+i. t+i ==k) is the probability that the weather type of the data sample at time t+i belongs to the kth class, h k,t+i Let be the probability weight of the weather type belonging to the kth class for the data sample at time t+i.

9. A computer device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the deep learning model parsing method for predicting the output of new energy power generation as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning model parsing method for predicting the output of new energy power generation as described in any one of claims 1-7.

Images