CN116700011A - Fractional calculus energy reduction guiding method for enhanced depth transducer-attribute integrated prediction - Google Patents

Abstract

The invention provides a fractional order calculus energy reduction guiding method for enhanced depth transducer-Attention integrated prediction. The method considers the energy consumption of the comprehensive energy system, takes the energy consumption influencing factors and the energy production as inputs, and outputs the optimal energy reduction guiding signals. The transform-Attention network and the high-efficiency time sequence prediction network combined with the time sequence Attention unit in the method can solve the problem of prediction of the reference energy consumption and are used for outputting the optimal prediction result; the fractional order random dynamic calculus controller in the method can obtain the optimal energy-reducing guide signal through the predicted energy consumption and energy production. The method can reduce the overall energy consumption of the comprehensive energy system and improve the stability of the comprehensive energy system.

Description

A Fractional Differential for Enhancing Deep Transformer-Attention Integrated Prediction Integral Degradation Guidance Method

technical field

The invention belongs to the energy control technology of electric power system, the field of artificial intelligence and the application field of calculus in mathematics application, relates to a control method of artificial intelligence and comprehensive energy system, and is applicable to the long-term energy reduction guidance of the comprehensive energy system.

Background technique

The patent application No. 2020108660490 filed on 2020.08.25 is titled "A Long-term Price Guidance Method for Multi-Group Distributed Flexible Energy Service Providers", which proposes a long-term dynamic game strategy under the condition of incomplete information. The long-term dynamic game among them does not include the influencing factors of the research object into the control strategy. 2022.01.18 The patent application No. 2021111974623 is titled "A Long-term Price Guidance Method Dynamic Differentiation Method for Dynamic Differential Control of Flexible Energy Hybrid Networks" only considers the case of random dynamic differential in integer order differential, but energy consumption and time There is a non-linear relationship between and has non-Markov properties, integer order stochastic dynamic differential has limitations in describing the relationship of non-Markov properties. The patent application No. 2022112767254 filed on 2022.10.18 is entitled "A Fractional Long-term Price Guidance Method for Enhancing Deep Attention Bidirectional Prediction", which uses fractional stochastic dynamic differentiation for energy consumption guidance of electric vehicles, but this method only It is limited to the single object of electric vehicles and does not take into account the entire energy system. The energy consumption of a single electric vehicle has very little impact on the entire energy system, and lacks energy management for the entire energy system.

Therefore, this paper proposes a fractional calculus energy reduction guidance method that enhances deep Transformer-Attention integrated prediction, which can consider the influence of energy consumption state, energy consumption coefficient, season, temperature and access rules on energy consumption; this method can Give full play to the regulating function of energy consumption in the integrated energy system, and solve the problem of energy supply and energy consumption imbalance in the integrated energy system; this method uses the Transformer-Attention network and the efficient timing prediction network combined with the temporal attention unit to analyze the user's The prediction of the reference energy consumption can solve the problem of predicting the user's energy consumption; this method introduces the fractional order integral of the Wiener process into the controller, which can better describe the random fluctuation of the noise signal and the random disturbance of the system; the method From the perspective of the integrated energy system, guide the system energy through the energy reduction guidance signal, and reduce energy consumption while satisfying the experience of the energy consumption side; in the long run, this method can reduce the energy consumption of the integrated energy system and improve the comprehensive energy system. stability.

Contents of the invention

The present invention proposes a fractional-order calculus energy reduction guidance method that enhances deep Transformer-Attention integrated prediction. Combined, it is used for long-term energy reduction guidance of the integrated energy system, which has the functions of improving energy utilization, improving the stability of the integrated energy system, promoting the integration of renewable energy into the integrated energy system, and reducing energy consumption; The steps are:

Step (1): Establish an operational framework for energy reduction guidance of the integrated energy system; the integrated energy system obtains energy production from power plants, boilers, and fossil fuels, and collects the expected energy consumption from the energy consumption side. The production volume is compared with the expected energy consumption, and the optimal energy reduction guidance signal is found through the energy reduction guidance method. Then, the integrated energy system sends the energy reduction guidance signal to the energy consumption side to guide industrial, commercial, and residential energy consumption. Reduce energy consumption;

Step (2): Propose a fractional-order calculus control that enhances deep Transformer-Attention integrated prediction, and predict the benchmark energy consumption of the integrated energy system through the Transformer-Attention network and the efficient timing prediction network combined with timing attention units;

First, data preprocessing is carried out, and the processed data is subjected to feature extraction; then, the processed feature data are respectively predicted by the Transformer-Attention network method and the efficient time series prediction network combined with the time series attention unit, and the prediction After the results are selected, the predicted baseline energy consumption is obtained;

The Transformer-Attention network is a deep network architecture with multi-head attention as the basic computing unit, which can achieve a balance between long-term dependence and low time complexity; by proposing the ProbSparse self-attention mechanism, self-attention distillation mechanism, and generative decoder to improve prediction accuracy;

Transformer-Attention network consists of input layer, encoder, decoder, fully connected layer and output layer;

Firstly, all the historical consumption of electric energy, thermal energy and fuel is input into the encoder for encoding to obtain a mapping sequence, and then the target value to be predicted in the long sequence is filled with zeros, and input together with the mapping sequence obtained by the encoder to In the decoder, the required predicted output elements are directly generated; the encoder shapes the t-th input sequence χ _t into a matrix:

In the formula, refers to the coding matrix; t refers to time t; /> refers to a set of real matrixes whose size is L _x ×d; refers to a real number; L _x refers to the length of the sequence; d is the input dimension;

The ProbSparse self-attention mechanism used by the Transformer-Attention network is different from the standard self-attention mechanism. The scaled dot product pair used by the standard self-attention mechanism is:

In the formula, A(Q, K, V) refers to the standard self-attention mechanism value; Q is the query vector; K refers to the queried vector; V is the content vector; KT is the transpose of the queried vector; Softmax( ) is a normalized exponential function; L _Q is the length of Q; L _K is the length of K; refers to the set of real matrix whose size is L _Q ×d;/> refers to a set of real matrixes whose size is L _K ×d;

The ProbSparse self-attention mechanism defines the scaled dot product pair of the standard attention mechanism as a probabilistic kernel smoother as

In the formula, A _i (q _i , K, V) refers to the self-attention mechanism value in the form of probability; i, j, l refers to the number of rows; q _i refers to the i-th row of Q; k _j refers to the jth row; k _l refers to the lth row of K; v _j refers to the jth row of V; k() refers to the probability function; p(k _j |q _i ) refers to the condition of k _j under q _i Probability distributions; is the sum of the conditional probabilities;

The asymmetric exponential kernel adopted by the probability function k(q _i , k _j ) in the kernel smoother is:

In the formula, is the transpose of k _j ;

The uniform distribution that the probability distribution of the query vector satisfies is:

In the formula, q(k _j |q _i ) refers to the uniform distribution of the query vector;

If p(k _j | q _i ) is close to the uniform distribution q(k _{j |} The important part of the sequence, the similarity measured by Kullback-Leibler divergence is:

where KL(q||p) refers to the Kullback-Leibler divergence between distribution p and q; refers to the transposition of k _l ; 1n() is a logarithmic function with base e;

Remove the lnL _K constant in the Kullback-Leibler divergence, and define the sparsity measure of the i-th query vector as:

In the formula, M(q _i , K) refers to the sparsity measure of the i-th query vector;

In the standard self-attention mechanism, the sparsity measurement obtains the attention value of the ProbSparse self-attention mechanism:

In the formula, A _s (Q, K, V) refers to the attention value of the ProbSparse self-attention mechanism; refers to a sparse matrix of the same size as Q, containing only the sparsity measure M(q _i , K);

By calculating the ProbSparse self-attention value of each element in each encoding matrix, the distillation process of the attention value is to extract the attention value of the ProbSparse self-attention mechanism, and give the attention value a better feature privilege with dominant features, and Generate focused self-attention feature maps in the next layer; Going forward from the nth layer to the (n+1) layer through the distillation process is:

In the formula, refers to /> Self-attention feature map for layer n; /> refers to /> The self-attention feature map of the (n+1)th layer;

MaxPool() refers to the maximum pooling function; ELU() refers to the activation function; Conv1d() refers to using the activation function to perform a 1-dimensional convolution filter in the time dimension; [] _AB refers to the multi-head ProbSparse self-attention and attention basic operations in the block;

The generative decoder is stacked by two identical multi-head attention layers, and the generative prediction can effectively alleviate the problem of slowing down in long-term prediction; the input vector to the decoder is:

In the formula, is a vector pointing to the input of the decoder; /> refers to the initial word segmentation vector; Refers to the placeholder vector of the target sequence, each element is 0; Ltoken refers to the length of the initial word segmentation vector; L _y refers to the length of the placeholder vector of the target sequence; Concat() refers to the splicing operation function;

by putting and /> After the decoding operation, finally pass the vector through the fully connected layer to obtain the final energy consumption that needs to be predicted:

In the formula, Refers to the energy consumption predicted by the Transformer-Attention network; /> It refers to the set of real matrix whose size is dy composed of the output value; dy refers to the dimension of the output data; /> refers to the o-th vector that constitutes the target output;

The efficient timing prediction network combined with timing attention unit does not use recurrent neural network, but uses attention mechanism to process time evolution in parallel; the efficient timing prediction network combining timing attention unit decomposes timing attention into two parts: Static attention and dynamic attention; static attention uses small core depth convolution and expanded convolution to achieve a large receptive field, thereby capturing the long-term dependency of the sequence; dynamic attention uses the difference in attention between time series to learn time series weights , so as to capture the change trend between sequences; the efficient temporal prediction network combined with temporal attention unit uses a differential dispersion regularization method to optimize the loss function of temporal prediction learning; the differential dispersion regularization method combines the predicted value with the real The difference between the values is converted into a probability distribution, and the Kullback-Leibler divergence between them is calculated, so that the efficient sequence prediction network combined with the sequence attention unit can learn the inherent change law in the sequence; the history of electric energy, thermal energy and fuel The consumption input into the input matrix of the efficient temporal prediction network combined with temporal attention unit is:

In the formula, T refers to the length of the input time series; refers to a set of real matrixes of size T;

The predicted value of the neural network map is:

In the formula, refers to the predicted value of the neural network model mapping; /> is a neural network model;

The forward difference between the predicted value and the real value mapped by the neural network model is:

In the formula, is the forward difference of the predicted value of the neural network map; /> refers to the i+1 predicted value mapped by the neural network model; /> refers to the i-th predicted value mapped by the neural network model; /> is the forward difference of the true value; refers to the true value /> i+1th data; /> refers to the true value /> i-th data;

The forward difference is transformed into a probability through the Softmax() function:

In the formula, σ() refers to the probability distribution function; refers to dynamic attention; /> refers to static attention; τ refers to the temperature coefficient; exp() is an exponential function with e as the base;

Calculate the probability distribution by and /> The Kullback-Leibler divergence between the micro-dispersion regularization function is:

In the formula, Refers to the microdispersion regularization function; T' is the length of the time series to be predicted;

An efficient temporal prediction network combined with a temporal attention unit is trained end-to-end in a fully unsupervised manner, and the loss function for evaluating variance composed of a mean squared error loss and a differential divergence regularization weighted by a constant λ is:

In the formula, is the loss function to evaluate the difference; λ is a constant;

The weight parameters of the efficient time series prediction network combined with the time series attention unit can be solved through the loss function as follows:

In the formula, Θ ^* is the value of the weight parameter obtained by solving; argmin is the solution corresponding to the minimum value of the objective function; Θ is a weight parameter;

The efficient temporal prediction network combined with temporal attention unit is to predict the subsequent T′ from time t+1, and the efficient temporal prediction network combined with temporal attention unit can learn from The mapping of , the energy consumption obtained by combining the efficient temporal prediction network with temporal attention unit is:

In the formula, refers to the energy consumption predicted by the efficient temporal prediction network combined with temporal attention units;/> refers to the set of real number matrices whose size is T′;

Step (3): Use the fractional-order calculus control predicted by the enhanced deep Transformer-Attention integration for energy reduction guidance of the integrated energy system; input the predicted baseline energy consumption into the fractional-order stochastic dynamic calculus controller, and the fractional-order The stochastic dynamic calculus controller outputs energy-reduction guiding signals;

The energy consumption of the integrated energy system includes power consumption, thermal energy consumption and fuel consumption. The energy consumption predicted by using the Transformer-Attention network and the efficient time-series prediction network combined with the time-series attention unit is obtained through the reference energy consumption prediction function. The amount is:

In the formula, D _t refers to the predicted baseline energy consumption; forecast() refers to the baseline energy consumption prediction function;

The energy consumption state differential output by the fractional order stochastic dynamic calculus controller is:

In the formula, α refers to the order of fractional calculus; S _t refers to the energy consumption state; d ^α S _t refers to the fractional order differential of the energy consumption state; refers to the amount of energy obtained by the integrated energy system; Pt refers to the energy consumption predicted by the fractional order stochastic dynamic calculus controller; d ^α t refers to the fractional order differential with respect to time; N _noise refers to the noise intensity; /> refers to the fractional order integration of the Wiener process; W _t refers to the Wiener process;

The comprehensive energy system includes the energy generation side and the energy consumption side. The energy consumption is affected by the change of the energy consumption output by the fractional order stochastic dynamic calculus controller as follows:

In the formula, refers to the variation of energy consumption; γ() refers to the logic function; l ₁ , l ₂ , l ₃ , l ₄ and l ₅ refer to the energy consumption state function and energy consumption coefficient function in the logic function γ() respectively , seasonal function, air temperature function and access rule function; Sta() refers to the energy consumption state function; δ is the parameter of the energy consumption state function Sta(); Coe ₍ ) refers to the energy consumption coefficient function; Energy consumption coefficient; β refers to the parameters of the energy consumption coefficient function Coe(); Wea() refers to the seasonal function; w _t refers to the seasonal situation; Tem() refers to the temperature function; h _t refers to the temperature; Rul() is refers to the access rule function; u _t refers to the access rule; θ refers to the parameter of energy consumption;

The predicted electricity load obtained by the fractional order stochastic dynamic differential controller is:

In the formula, p _energy refers to the proportion of renewable energy; is a sign function;

The symbolic function S() is:

The logic function γ() is:

In the formula, l is the parameter of the logic function γ();

Step (4): Considering the dynamic energy consumption changes caused by the energy consumption state, energy consumption coefficient, season, air temperature and access rules, a fractional order stochastic dynamic calculus controller generates an energy reduction guidance signal;

The influencing factors of energy consumption, the predicted baseline energy consumption and the energy provided by the integrated energy system are used as the input variables of the fractional-order stochastic dynamic calculus controller, and the energy reduction guide signal is used as the output variable;

The energy consumption state function Sta() is:

Sta(S _t , δ ₁ , δ ₂ , δ ₃ , δ ₄ )=(1-2S _t +δ ₁ ×[1-(2S _t -1) ² ]}×[δ ₂ +δ ₃ ×(2S _t -1) ² +δ ₄ ×(2S _t -1) ⁶ ] (26)

In the formula, δ ₁ , δ ₂ , δ ₃ and δ ₄ refer to the degree of skewness, constant term of change, quadratic term of change and sixth order of change in the energy consumption state function Sta(), respectively coefficient;

The energy consumption coefficient function Coe() is:

In the formula, TNz refers to the total number of splines; z refers to the zth spline; Iz() refers to the I-spline function;

The seasonal function Wea() is:

In the formula, sin() refers to the sine function;

The temperature function Tem() is:

Tem(h _t )＝0.6 exp(h _t )+8h _t (29)

The access rule function Rul() is:

The predicted baseline energy consumption and energy supply are used as the input variables of the fractional-order stochastic dynamic calculus controller, and the predicted energy consumption is output through the fractional-order stochastic dynamic differential equation, and then the optimal energy reduction guide signal is solved by using the objective function; The function of the fractional-order stochastic dynamic calculus controller to solve the degraded guiding signal is:

In the formula, period refers to the forecast period; Pre(D _t , c _t ) refers to the predicted energy consumption function with the energy consumption coefficient c _t as a variable;

Step (5): Apply the energy reduction guidance signal to the integrated energy system, guide the energy consumption side to use energy, improve energy utilization, strengthen the stability of the integrated energy system, promote the integration of renewable energy, and reduce the energy consumption of the integrated energy system. energy consumption.

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

(1) The present invention utilizes the data preprocessing function, the Transformer-Attention network and the efficient time series prediction network combined with the time series attention unit to predict the user's baseline energy consumption, and at the same time introduces the fractional order integration of the Wiener process into the fractional order stochastic dynamics In the calculus controller, the accuracy of the degraded pilot signal can be improved.

(2) From the perspective of the comprehensive energy system, the present invention considers the influence of energy consumption state, energy consumption coefficient, season, temperature and access rules on energy consumption, guides the energy consumption side to consume energy through the energy reduction guidance signal, and reduces the comprehensive energy consumption. The energy consumption of the system.

(3) Compared with the patent application No. 2020108660490 filed on August 25, 2020, the present invention is entitled "A Long-term Price Guidance Method for Multi-Group Distributed Flexible Energy Service Providers", which not only considers the dynamics between energy production and energy consumption The game also takes the influence factors of energy consumption into the energy reduction guidance of the comprehensive energy system.

(4) Compared with the patent application No. 2021111974623 filed on 2022.01.18, the present invention is entitled "A Dynamic Differentiation Method for Long-term Price Guidance Method for Dynamic Differential Control of Flexible Energy Hybrid Network", which adds fractional order stochastic dynamic differentiation to the control In the detector, the nonlinear relationship and non-Markovian nature between energy consumption and time can be better described.

(5) Compared with the patent application No. 2022112767254 filed on 2022.10.18, the present invention is entitled "A Fractional Long-term Price Guidance Method for Enhancing Deep Attention Bidirectional Prediction", which expands the research object from a single electric vehicle to a comprehensive In the energy system, at the same time, the fractional order integral of the Wiener process is added to the controller, which can better describe the random fluctuation of the noise signal and the random disturbance of the system, and improve the accuracy of the energy consumption guidance signal.

Description of drawings

Fig. 1 is a control framework diagram of the fractional-order calculus energy reduction guidance method for enhanced deep Transformer-Attention integrated prediction of the method of the present invention.

Figure 2 is the Transformer-Attention network of the method of the present invention.

Fig. 3 is an efficient temporal prediction network combined with a temporal attention unit of the method of the present invention.

Detailed ways

The present invention proposes a fractional-order calculus energy reduction guidance method that enhances deep Transformer-Attention integrated prediction. The detailed description is as follows in conjunction with the accompanying drawings:

Fig. 1 is a control framework diagram of the fractional-order calculus energy reduction guidance method for enhanced deep Transformer-Attention integrated prediction of the method of the present invention. First, the raw energy consumption data obtained from industry, commerce, and residences are preprocessed. Then, the preprocessed data is predicted through the Transformer-Attention network and the efficient time-series prediction network combined with the time-series attention unit to determine the predicted baseline energy consumption. Finally, input the predicted baseline energy consumption into the fractional-order stochastic dynamic calculus controller, combine energy consumption factors and energy production, and output the optimal energy reduction guidance signal to guide industrial, commercial, and residential energy consumption.

Figure 2 is the Transformer-Attention network of the method of the present invention. First the encoder receives a large number of long sequence inputs. The Transformer-Attention network replaces standard self-attention with ProbeSparse self-attention. Then, the distillation operation is performed on the multi-head attention. The attention after the distillation operation is a pyramid dependency. Through layer-by-layer distillation, the dominant attention is extracted to the decoder, and the network size is greatly reduced. Finally, the decoder takes long sequence inputs, zero-pads target elements, computes a weighted attention composition of feature maps, and immediately predicts output elements in a generative style.

Fig. 3 is an efficient temporal prediction network combined with a temporal attention unit of the method of the present invention. The overall structure of an efficient temporal prediction network combined with temporal attention units is data input, encoder, temporal attention unit, decoder, and data output. In the temporal attention unit, firstly, the static attention generated by the actual value is converted by small kernel depth convolution, dilated depth convolution and 1*1 convolution to obtain the first output vector. Then, the dynamic attention of the neural network map is passed through the average pooling layer and the fully connected layer to obtain the second output vector. Finally, the two output vectors are combined and sent to the decoder, which generates the final data output.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process conversion made by using the description and drawings of the present invention, or directly or indirectly used in other relevant Technical fields are all included in the scope of patent protection of the present invention in the same way.

Claims (1)

1. The fractional calculus energy reduction guiding method for enhancing depth conversion-Attention integrated prediction is characterized by combining a conversion-Attention network, a high-efficiency time sequence prediction network combined with a time sequence Attention unit and a fractional random dynamic calculus controller, is used for long-term energy reduction guiding of a comprehensive energy system, and has the functions of improving the energy utilization rate, improving the stability of the comprehensive energy system, promoting renewable energy to be integrated into the comprehensive energy system and reducing energy consumption; the method comprises the following steps in the using process:

step (1): establishing an operation frame for energy reduction guidance of the comprehensive energy system; the comprehensive energy system obtains energy production from power plants, boilers and fossil fuels, collects expected required energy consumption from an energy consumption side, compares the energy production with the expected required energy consumption, finds an optimal energy-reducing guide signal through an energy-reducing guide method, and then sends the energy-reducing guide signal to the energy consumption side to guide energy consumption of industries, businesses and residences and reduce the energy consumption;

step (2): providing fractional calculus control for enhancing depth transducer-Attention integrated prediction, and predicting the reference energy consumption of the comprehensive energy system through a transducer-Attention network and a high-efficiency time sequence prediction network combined with a time sequence Attention unit;

firstly, preprocessing data, and extracting features of the processed data; then, respectively predicting the processed characteristic data by using a transducer-Attention network method and a high-efficiency time sequence prediction network combined with a time sequence Attention unit, and selecting a prediction result to obtain a predicted reference energy consumption;

the transform-Attention network is a deep network architecture taking multi-head Attention as a basic operation unit, and can acquire balance between long-term dependence acquisition and low time complexity acquisition; the prediction precision is improved by proposing a ProbSparse self-attention mechanism, a self-attention distillation mechanism and a generative decoder;

the transducer-attribute network consists of an input layer, an encoder, a decoder, a full-connection layer and an output layer;

firstly, all the historical consumption of electric energy, heat energy and fuel are input into an encoder to be encoded to obtain a mapping sequence, then target values to be predicted in a long sequence are filled to be zero, the target values and the mapping sequence obtained by the encoder are input into a decoder together, and prediction output elements to be obtained are directly generated; the encoder inputs the t-th input sequence χ _t Molding into a matrix:

in the method, in the process of the invention,refers to a coding matrix; t is the time t; />Refers to the size L _x X d set of real matrices; />Refers to real numbers; l (L) _x Refers to the length of the sequence; d is the input dimension;

the probspark self-Attention mechanism employed by the transform-Attention network differs from the standard self-Attention mechanism, which employs scaled dot product pairs of:

wherein A (Q, K, V) is a self-attention mechanism value indicative of the index; q is a query vector; k refers to the vector being queried; v is the content vector; k (K) ^T Is a transpose of the queried vector; softmax () is a normalized exponential function; l (L) _Q Is the length of Q; l (L) _K Is the length of K;refers to the size L _Q X d set of real matrices; />Refers to the size L _K X d set of real matrices;

the probspark self-attention mechanism defines the scaled dot product pair of the standard attention mechanism as a kernel smoother in probabilistic form as:

wherein A is _i (q _i K, V) refers to the self-attention mechanism value in the form of probability; i. j and l refer to the number of rows; q _i Refers to row i of Q; k (k) _j Refers to row j of K; k (k) _l Refers to the first row of K; v _j Refers to row j of V; k () refers to a probability function; p (k) _j |q _i ) Refers to k _j At q _i The following conditional probability distribution;refers to the sum of conditional probabilities;

probability function k (q in kernel smoother _i ,k _j ) The asymmetric index kernel used is:

in the method, in the process of the invention,is k _j Is a transpose of (2);

the probability distribution of the query vector satisfies the uniform distribution:

wherein q (k) _j |q _i ) Refers to the uniform distribution of query vectors;

if p (k) _j |q _i ) Near uniform distribution q (k) _j |q _i ) The self-attention becomes the value V, which is redundant for the prediction output, so the similarity between the distributions p and q is used to distinguish important parts of the sequence, as measured by Kullback-Leibler divergence:

wherein KL (q||p) refers to the Kullback-Leibler divergence between the distributions p and q;refers to k _l Is a transpose of (2); ln () is a logarithmic function based on e;

removal of lnL in the Kullback-Leibler divergence _K This constant, the sparsity measure for the ith query vector is defined as:

wherein M (q _i K) refers to the sparsity measure of the ith query vector;

the sparsity measure in the standard self-attention mechanism yields the attention value of the probspark self-attention mechanism as:

wherein A is _s (Q, K, V) refers to the attention value of the ProbSparse self-attention mechanism;refers to a sparse matrix of the same size as Q, comprising only sparsity metrics M (Q _i ,K)；

By calculating the ProbSparse self-attention value of each element in each coding matrix, the distillation process of the attention value is to extract the attention value of the ProbSparse self-attention mechanism, endow the attention value with better feature privileges with dominant features, and generate a focused self-attention feature map at the next layer;the forward entry into the (n+1) layer from the n-th layer by distillation is:

in the method, in the process of the invention,refers to->A layer n self-care feature map; />Refers to->A self-care feature map of layer (n+1); maxPool () refers to the max pooling function; ELU () refers to an activation function; conv1d () refers to performing a 1-dimensional convolution filter in the time dimension using an activation function; [] _AB Refers to the basic operations in the self-attention and attention block that contain multiple heads probspark;

the generating type decoder is formed by stacking 2 identical multi-head attention layers, and the generating type prediction can effectively relieve the problem of speed reduction in long-time prediction; the vector input to the decoder is:

in the method, in the process of the invention,is a vector pointing to the decoder input; />The initial word segmentation vector is referred to; />Refers to a placeholder vector for the target sequence, each element being 0; l (L) _token Refers to the length of the initial segmentation vector; l (L) _y Refers to the length of the placeholder vector of the target sequence; concat () refers to a splice operation function;

by combiningAnd->Finally, the vector is passed through the full connection layer after decoding operation to obtain the final energy consumption to be predicted, which is:

in the method, in the process of the invention,the energy consumption is predicted by a transducer-attribute network; />Refers to the output value with the size d _y A real number matrix set; d, d _y Refers to the dimension of the output data; />Refers to the o-th vector constituting the target output;

the efficient timing prediction network in combination with the timing attention unit does not use a recurrent neural network, but rather uses an attention mechanism to parallelize the processing of the time evolution; an efficient timing prediction network in combination with a timing attention unit breaks the timing attention down into two parts: static attention and dynamic attention; static attention uses small core depth convolution and dilation convolution to achieve a large receptive field, capturing long-term dependencies of sequences; the dynamic attention learns the time sequence weight by utilizing the difference of the time sequence attention, thereby capturing the variation trend among the sequences; the high-efficiency time sequence prediction network combined with the time sequence attention unit utilizes a differential divergence regularization method for optimizing a loss function of time sequence prediction learning; the differential divergence regularization method converts the difference between the predicted value and the true value into probability distribution, calculates the Kullback-Leibler divergence between the predicted value and the true value, and enables an efficient time sequence prediction network combined with a time sequence attention unit to learn the inherent change rule in the time sequence; the input matrix of the high-efficiency time sequence prediction network combined with the time sequence attention unit is that the historical consumption of electric energy, heat energy and fuel is input as follows:

wherein T is the length of the input time series;refers to a real matrix set of size T;

the predicted values for the neural network map are:

in the method, in the process of the invention,the predicted value of the neural network model mapping is referred to; />Is a neural network model;

the predicted value and the true value mapped by the neural network model are subjected to forward difference as follows:

in the method, in the process of the invention,is the forward difference of the predicted values of the neural network map; />The (i+1) th predicted value mapped by the neural network model is referred to; />Refers to the ith predicted value mapped by the neural network model; />Refers to the forward difference of the true values; />Refers to the true value +.>Data i+1th; />Refers to the true value +.>Ith data;

the forward difference is converted into probability by Softmax () function as follows:

wherein σ () refers to a probability distribution function;refers to dynamic attention; />Refers to static attention; τ is the temperature coefficient; exp () is an exponential function based on e;

by calculating probability distributionAnd->The obtained micro-dispersion regularization function of the Kullback-Leibler dispersion is as follows:

in the method, in the process of the invention,refers to a micro-dispersion regularization function; t' is the length of the time series that needs to be predicted;

the efficient time sequence prediction network combined with the time sequence attention unit performs end-to-end training in a completely unsupervised manner, and the loss function of the evaluation difference consisting of the mean square error loss and the constant lambda weighted differential divergence regularization is as follows:

in the method, in the process of the invention,is a loss function that evaluates the difference; lambda is a constant;

the weight parameters of the high-efficiency time sequence prediction network combined with the time sequence attention unit can be solved through the loss function, and the weight parameters are as follows:

in the formula Θ ^* Is the value of the solved weight parameter; argmin is a solution corresponding to the minimum value of the solving objective function; Θ is a weight parameter;

the efficient time sequence prediction network combined with the time sequence attention unit predicts the following T' from the time t+1, and the efficient time sequence prediction network combined with the time sequence attention unit can learn the slaveThe energy consumption obtained by the high-efficiency time sequence prediction network combined with the time sequence attention unit is as follows:

in the method, in the process of the invention,refers to the energy consumption predicted by the network through the efficient time sequence prediction combined with the time sequence attention unit;refers to a real matrix set of size T';

step (3): fractional order calculus control of enhanced depth transducer-attribute integrated prediction is used for energy reduction guidance of a comprehensive energy system; inputting the predicted reference energy consumption into a fractional order random dynamic calculus controller, and outputting an energy reduction guide signal by the fractional order random dynamic calculus controller;

the energy consumption of the comprehensive energy system comprises electric power consumption, heat energy consumption and fuel consumption, and the reference energy consumption obtained by using the energy consumption predicted by the converter-Attention network and the high-efficiency time sequence prediction network combined with the time sequence Attention unit through the reference energy consumption prediction function is as follows:

wherein D is _t Refers to a predicted reference energy consumption; forecast () refers to a reference energy consumption amount prediction function;

the energy consumption state differentiation output by the fractional order random dynamic calculus controller is as follows:

wherein, alpha refers to the order of fractional calculus; s is S _t Refers to the energy consumption state; d, d ^α S _t Refers to fractional differentiation of the energy consumption state;refers to the energy obtained by the comprehensive energy system; p (P) _t The energy consumption predicted by the fractional order random dynamic calculus controller is referred to; d, d ^α t refers to fractional differentiation over time; n (N) _noise Refers to noise intensity; />Refers to fractional order integration of the wiener process; w (W) _t Refers to the wiener process;

the comprehensive energy system comprises an energy generation side and an energy consumption side, wherein the energy consumption is output by the utilization fractional order random dynamic calculus controller, and the change of the energy consumption is as follows:

in the method, in the process of the invention,refers to the amount of change in energy consumption; gamma () refers to a logical function; l (L) ₁ 、l ₂ 、l ₃ 、l ₄ And l ₅ The energy consumption state function, the energy consumption coefficient function, the seasonal function, the air temperature function and the admission rule function in the logic function gamma (); sta () refers to an energy consumption state function; delta is a parameter of the energy consumption state function Sta (); coe () refers to an energy consumption coefficient function; c _t Refers to the energy consumption coefficient; beta refers to a parameter of an energy consumption coefficient function Coe (); wea () refers to a seasonal function; w (w) _t Refers to the seasonal situation; tem () refers to an air temperature function; h is a _t Refers to air temperature; rul () refers to an admission rule function; u (u) _t Refers to an admission rule; θ refers to a parameter of energy consumption;

the predicted electricity load is obtained through a fractional order random dynamic differential controller as follows:

wherein p is _energy Refers to the proportion of renewable energy sources;refers to a sign function;

the sign function S () is:

the logical function γ () is:

where l is a argument of a logical function γ ();

step (4): taking into account dynamic energy consumption changes caused by energy consumption states, energy consumption coefficients, seasons, air temperatures and admission rules, generating a descending energy guiding signal by a fractional order random dynamic calculus controller;

the influence factors of energy consumption, the predicted reference energy consumption and the energy provided by the comprehensive energy system are used as input variables of the fractional order random dynamic calculus controller, and the energy reduction guide signal is used as output variable;

the energy consumption state function Sta () is:

Sta(S _t ,δ ₁ ,δ ₂ ，δ ₃ ,δ ₄ )＝{1-2S _t +δ ₁ ×[1-(2S _t -1) ^z l}×[δ ₂ +δ ₃ ×(2S _t -1) ^z +δ ₄ ×(2S _t -1) ⁶ ] (26)

in delta ₁ 、δ ₂ 、δ ₃ And delta ₄ The energy consumption state function Sta () is a coefficient for controlling the deflection degree, the variable constant term, the variable quadratic term and the variable sixth term of the energy consumption state;

the energy consumption coefficient function Coe () is:

in TN _z Refers to the total number of splines; z refers to the z-th spline; i _z () Refers to I spline functions;

the seasonal function Wea () is:

where sin () refers to a sine function;

the air temperature function Tem () is:

Tem(h _t )＝0.6exp(h _t )+8h _t (29)

the admission rule function Rul () is:

the predicted reference energy consumption and the energy supply are used as input variables of a fractional order random dynamic calculus controller, the predicted energy consumption is output through a fractional order random dynamic derivative equation, and then an optimal energy reduction guide signal is solved by utilizing an objective function; the function of solving the energy-reducing guide signal by using the fractional order random dynamic calculus controller is as follows:

wherein period refers to a prediction period; pre (D) _t ,c _t ) Is expressed by the energy consumption coefficient c _t A predicted energy consumption function for the variable;

step (5): the energy-reducing guide signal is applied to the comprehensive energy system, the energy consumption side is guided to use energy, the energy utilization rate is improved, the stability of the comprehensive energy system is enhanced, the integration of renewable energy sources is promoted, and the energy consumption of the comprehensive energy system is reduced.