CN117674091A

CN117674091A - Photovoltaic output power prediction method and related device based on dual-attention network

Info

Publication number: CN117674091A
Application number: CN202311596123.1A
Authority: CN
Inventors: 窦猛汉; 请求不公布姓名
Original assignee: Benyuan Quantum Computing Technology Hefei Co ltd
Current assignee: Benyuan Quantum Computing Technology Hefei Co ltd
Priority date: 2023-11-23
Filing date: 2023-11-23
Publication date: 2024-03-08

Abstract

The invention discloses a photovoltaic output power prediction method and a related device based on a dual-attention network, which are applied to the field of quantum computation and comprise the following steps: based on photovoltaic data of a plurality of time nodes before the time node to be predicted, predicting by using a quantum long-short-time memory network to obtain a hidden state of photovoltaic output power of the time node to be predicted; processing the hidden state by using a dual-attention network to obtain a first weight of the hidden state in a time dimension and a second weight of the hidden state in a characteristic dimension; constructing a target output matrix based on the association relation among the first weight, the second weight and the hidden state; and carrying out full-connection decision processing on the time dimension and the characteristic dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted. The accuracy of photovoltaic output power prediction is further improved by analyzing the change condition of a plurality of time nodes of the photovoltaic output power in the time dimension and the influence of characteristics of different dimensions on the output power.

Description

Photovoltaic output power prediction method and related device based on dual-attention network

Technical Field

The invention belongs to the technical field of quantum computation, and particularly relates to a photovoltaic output power prediction method and a related device based on a dual-attention network.

Background

The photovoltaic output power represents the electric energy generated by the photovoltaic panel in unit time, and is an important index for measuring the output electric energy of the photovoltaic power generation system, and the output power directly influences the actual power generation capacity and economic benefit of the photovoltaic panel. The photovoltaic output power of the photovoltaic power station is affected by factors such as meteorological fluctuation, and when the photovoltaic power station is connected to a power grid, the output power is unstable and is not beneficial to the dispatching management of the power grid power. Therefore, accurate prediction of photovoltaic output power is crucial to scheduling management of electric power, and currently, output power in a future period of time is predicted based on historical photovoltaic data mainly by using a neural network model.

The LSTM (Long Short Term Memory long and short term memory) network is an improvement and optimization of the cyclic neural network, and can learn long-term information through a uniquely designed gating system, so that the long-term memory of data input into the neural network is realized, and the LSTM network can be used for processing more complex time sequence problems. Correspondingly, the quantum computer is a physical device for carrying out high-speed mathematical and logical operation, storing and processing quantum information according to the law of quantum mechanics, and has the capability of processing mathematical problems more efficiently than a classical computer. In order to better handle high-dimensional and high-complexity historical photovoltaic data, a quantum long-short-term memory network can be constructed by replacing a classical neural network in an LSTM network with a variable component sub-line.

However, the quantum long-short-term memory network still predicts future photovoltaic output power mainly based on photovoltaic data at the previous moment, and cannot fully understand the historical change condition of the photovoltaic output power and the importance degree of the influence of the characteristics of different dimensions on the photovoltaic output power. Therefore, there is a need to further improve the accuracy of photovoltaic output power predictions.

Disclosure of Invention

The invention aims to provide a photovoltaic output power prediction method based on a dual-attention network and a related device, aiming at further improving the accuracy of photovoltaic output power prediction.

One embodiment of the invention provides a photovoltaic output power prediction method based on a dual-attention network, which comprises the following steps:

based on photovoltaic data of a plurality of time nodes before a time node to be predicted, predicting by using a quantum long-short-time memory network to obtain a hidden state of photovoltaic output power of the time node to be predicted;

processing the hidden state by using a dual-attention network to obtain a first weight and a second weight; the first weight is the weight of the hidden state in the time dimension, and the second weight is the weight of the hidden state in the characteristic dimension;

Constructing a target output matrix based on the association relation among the first weight, the second weight and the hidden state;

and carrying out full-connection decision processing on the time dimension and full-connection decision processing on the characteristic dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted.

Optionally, the dual attention network includes a temporal attention network and a feature attention network;

the processing the hidden state by using the dual-attention network to obtain a first weight and a second weight includes:

extracting time weight components on each time node in the hidden state by using the time attention network to obtain a first weight;

and extracting feature weight components on each feature dimension in the hidden state by using the feature attention network to obtain a second weight.

Optionally, the constructing the target output matrix based on the association relationship between the first weight, the second weight and the hidden state includes:

constructing a first output matrix based on the temporal weight component and the state components in the hidden state;

constructing a second output matrix based on the feature weight components and the state components in the hidden state; and calculating the sum of the first output matrix and the second output matrix as the target output matrix.

Optionally, the hidden state is a hidden feature matrix constructed based on hidden features of a plurality of time nodes, each column vector of the hidden feature matrix corresponds to a hidden feature of a time node, and each row vector corresponds to a hidden feature of a feature dimension;

the extracting, by using the time attention network, time weight components on each time node in the hidden state to obtain a first weight includes:

for each column vector in the hidden feature matrix, carrying out average pooling on elements in the column vector to obtain a one-dimensional time array corresponding to the hidden feature matrix;

performing convolution processing on the one-dimensional time array to obtain time weight components corresponding to each time node, and constructing the first weight based on the time weight components;

extracting feature weight components on each feature dimension in the hidden state by using the feature attention network to obtain a second weight, wherein the method comprises the following steps:

for each row of vectors in the hidden feature matrix, carrying out average pooling on elements in the row of vectors to obtain a one-dimensional feature array corresponding to the hidden feature matrix;

and carrying out convolution processing on the one-dimensional feature array to obtain feature weight components corresponding to each feature dimension, and constructing the second weight based on the feature weight components.

said constructing a first output matrix based on said temporal weight component and a state component in said hidden state, comprising:

calculating the product of each element in the hidden characteristic matrix and the time weight component of the corresponding time node to be used as the element in the first output matrix to obtain the first output matrix;

said constructing a second output matrix based on said feature weight components and state components in said hidden state, comprising:

and calculating the product of each element in the hidden characteristic matrix and the characteristic weight component of the corresponding characteristic dimension to be used as the element in the second output matrix to obtain the second output matrix.

Optionally, before the photovoltaic data of the time nodes before the time node to be predicted is based on the photovoltaic data of the time nodes, and the hidden state of the photovoltaic output power of the time node to be predicted is obtained by using the quantum long-short-time memory network, the method further includes:

Acquiring photovoltaic output power data and meteorological characteristic data of a plurality of time nodes before the time node to be predicted;

constructing a track matrix of the photovoltaic output power based on the photovoltaic output power data of the plurality of time nodes, and performing singular value decomposition on the track matrix;

performing feature grouping and diagonal averaging on the submatrices obtained by singular value decomposition to obtain a power trend sequence, a power cycle sequence and a noise sequence of the plurality of time nodes;

and carrying out normalization processing on the meteorological characteristic data, the photovoltaic output power data, the power trend sequence, the power cycle sequence and the noise sequence to obtain the photovoltaic data.

Optionally, the predicting, based on the photovoltaic data of a plurality of time nodes before the time node to be predicted, the hidden state of the photovoltaic output power of the time node to be predicted by using the quantum long-short time memory network includes:

vector splicing is carried out on the photovoltaic data of the time node and the hidden state corresponding to the time node according to a preset proportion aiming at each time node before the time node to be predicted, so as to obtain a corresponding spliced vector; wherein, the hidden state corresponding to the first time node is a preset value;

Inputting the spliced vector into the quantum long-short-time memory network, and operating the quantum long-short-time memory network to obtain a hidden state corresponding to the next time node of the time node;

and determining the hidden state of the photovoltaic output power of the time node to be predicted based on the hidden states corresponding to all the time nodes.

Optionally, the quantum long-short-time memory network includes a plurality of variable component sub-lines, each of the variable component sub-lines includes a coding line and a variable component line connected in sequence, the coding line includes a first single-quantum logic gate combination acting on each of the qubits, and the variable component line includes a second single-quantum logic gate combination acting on each of the qubits and a multiple-quantum logic gate acting on the plurality of the qubits, wherein:

the first single quantum logic gate combination comprises a cascaded H gate, a first RY gate, and a first RZ gate; wherein the rotation parameters of the first RY gate and the first RZ gate are determined based on the splice vector;

the second single quantum logic gate combination comprises a second RZ gate, a second RY gate and a third RZ gate which are cascaded, and the multiple quantum logic gate is a CRX gate; wherein the rotational parameters of the second RZ gate, the second RY gate, the third RZ gate, and the CRX gate are determined based on training.

Yet another embodiment of the present invention provides a dual-attention network-based photovoltaic output power prediction apparatus, the apparatus comprising:

the prediction module is used for predicting the hidden state of the photovoltaic output power of the time node to be predicted by utilizing the quantum long-short-time memory network based on the photovoltaic data of a plurality of time nodes before the time node to be predicted;

the weight module is used for processing the hidden state by utilizing a dual-attention network to obtain a first weight and a second weight; the first weight is the weight of the hidden state in the time dimension, and the second weight is the weight of the hidden state in the characteristic dimension;

the output module is used for constructing a target output matrix based on the association relation among the first weight, the second weight and the hidden state;

and the decision module is used for carrying out full-connection decision processing on the time dimension and full-connection decision processing on the characteristic dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted.

A further embodiment of the invention provides a storage medium having a computer program stored therein, wherein the computer program is arranged to perform the method of any of the preceding claims when run.

Yet another embodiment of the invention provides an electronic device comprising a memory having a computer program stored therein and a processor configured to run the computer program to perform the method described in any of the above.

Compared with the prior art, the photovoltaic output power prediction method and device based on the dual-attention network, the storage medium and the electronic device provided by the invention comprise the following steps: based on photovoltaic data of a plurality of time nodes before the time node to be predicted, predicting by using a quantum long-short-time memory network to obtain a hidden state of photovoltaic output power of the time node to be predicted; processing the hidden state by using a dual-attention network to obtain a first weight and a second weight, wherein the first weight is the weight of the hidden state in the time dimension, and the second weight is the weight of the hidden state in the characteristic dimension; constructing a target output matrix based on the association relation among the first weight, the second weight and the hidden state; and carrying out full-connection decision processing on the time dimension and full-connection decision processing on the characteristic dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted.

According to the quantum long-short-term memory network in the technical scheme, the complex nonlinear relation between the photovoltaic data is efficiently captured by utilizing quantum superposition and quantum entanglement characteristics of quantum states, and accurate modeling is carried out on the complex nonlinear relation, so that the hidden state of the photovoltaic output power of the time node to be predicted can be predicted based on the photovoltaic data of a plurality of time nodes before the time node to be predicted. And respectively extracting weight information in the time dimension and the characteristic dimension in the hidden state by utilizing a dual-attention network, and constructing a target output matrix by analyzing the change condition of a plurality of time nodes of the photovoltaic output power in the time dimension and the influence of the characteristics of different dimensions on the photovoltaic output power. And the photovoltaic output power prediction result obtained through the full-connection decision processing in the time dimension and the full-connection decision processing in the feature dimension has higher accuracy.

Drawings

FIG. 1 is a network block diagram of a photovoltaic output power prediction system provided by an embodiment of the present invention;

fig. 2 is a schematic flow chart of a photovoltaic output power prediction method based on a dual-attention network according to an embodiment of the present invention;

fig. 3 is a schematic flow chart of obtaining photovoltaic data according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a specific flow for obtaining a hidden state by quantum long-short-term memory network prediction according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a specific structure of a quantum long-short time memory network according to an embodiment of the present invention;

fig. 6 is a schematic diagram of a specific structure of a variable component sub-circuit according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of a specific flow for constructing a target output matrix according to an embodiment of the present invention;

fig. 8 is a schematic diagram of a photovoltaic output power prediction result provided by an embodiment of the present invention;

fig. 9 is a schematic structural diagram of a photovoltaic output power prediction device based on a dual-attention network according to an embodiment of the present invention;

fig. 10 is a schematic structural diagram of a computer device according to an embodiment of the present invention.

Detailed Description

The embodiments described below by referring to the drawings are illustrative only and are not to be construed as limiting the invention.

Fig. 1 is a network block diagram of a photovoltaic output power prediction system according to an embodiment of the present invention. The photovoltaic output power prediction system may include a network 110, a server 120, a wireless device 130, a client 140, a store 150, a classical computing unit 160, a quantum computing unit 170, and may also include additional memory, classical processors, quantum processors, and other devices not shown.

Network 110 is a medium used to provide a communication link between various devices and computers connected together within the photovoltaic output power prediction system, including but not limited to the internet, intranets, local area networks, mobile communication networks, and combinations thereof, and may be connected by wired, wireless communication links, or fiber optic cables, etc.

Server 120, wireless device 130, and client 140 are conventional data processing systems that may contain data and have applications or software tools that perform conventional computing processes. The client 140 may be a personal computer or a network computer, so the data may also be provided by the server 120. The wireless device 130 may be a smart phone, tablet, notebook, smart wearable device, or the like. The memory unit 150 may include a database 151 that may be configured to store data of qubit parameters, quantum logic gate parameters, quantum wires, quantum programs, and the like.

Classical computing unit 160 (quantum computing unit 170) may include classical processor 161 (quantum processor 171) for processing classical data (quantum data), which may be boot files, operating system images, and application 163 (application 173), and memory 162 (memory 172) for storing classical data (quantum data), which may be quantum algorithms compiled by application 163 (application 173) for implementing a dual-attention network-based photovoltaic output power prediction method provided in accordance with embodiments of the present invention.

Any data or information stored or generated in classical computing unit 160 (quantum computing unit 170) may also be configured to be stored or generated in another classical (quantum) processing system in a similar manner, as may any application program that it executes.

It should be noted that, the real quantum computer is a hybrid structure, and it includes at least two major parts in fig. 1: a classical calculation unit 160 responsible for performing classical calculations and controls; the quantum computing unit 170 is responsible for running a quantum program to realize quantum computing.

The classical computing unit 160 and the quantum computing unit 170 may be integrated in one device or may be distributed among two different devices. A first device, for example, comprising a classical computing unit 160 runs a classical computer operating system on which quantum application development tools and services are provided, and also the storage and network services required for quantum applications. The user develops the quantum program through a quantum application development tool and service thereon, and transmits the quantum program to a second device including the quantum computing unit 170 through a web service thereon. The second device runs a quantum computer operating system, the code of the quantum program is analyzed and compiled into an instruction which can be identified and executed by the quantum processor 170 through the quantum computer operating system, and the quantum processor 170 realizes a quantum algorithm corresponding to the quantum program according to the instruction.

The computation unit of the classical processor 161 in the classical computation unit 160 is a CMOS tube based on silicon chips, which is not limited by time and coherence, i.e. which is not limited by the time of use, which is available at any time. Furthermore, in silicon chips, the number of such computation units is also sufficient, the number of computation units in a classical processor 161 is now thousands, the number of computation units is sufficient and the CMOS pipe selectable computation logic is fixed, for example: and AND logic. When the CMOS tube is used for operation, a large number of CMOS tubes are combined with limited logic functions, so that the operation effect is realized.

The basic computational unit of quantum processor 171 in quantum computational unit 170 is a qubit, the input of which is limited by coherence and also by coherence time, i.e., the qubit is limited in terms of time of use and is not readily available. Full use of qubits within the usable lifetime of the qubits is a critical challenge for quantum computing. Furthermore, the number of qubits in a quantum computer is one of the representative indicators of the performance of the quantum computer, each of the qubits realizes a calculation function by a logic function configured as needed, whereas the logic function in the field of quantum calculation is diversified in view of the limited number of qubits, for example: hadamard gates (Hadamard gates, H gates), brix-gates (X gates), brix-Y gates (Y gates), brix-Z gates (Z gates), X gates, RY gates, RZ gates, CNOT gates, CR gates, issnap gates, toffoli gates, and the like. In quantum computation, the operation effect is realized by combining limited quantum bits with various logic function combinations.

Based on these differences, the design of classical logic functions acting on CMOS transistors and the design of quantum logic functions acting on qubits are significantly and essentially different; the classical logic function acts on the design of the CMOS tube without considering the individuality of the CMOS tube, such as the individuality identification and the position of the CMOS tube in the silicon chip, and the usable time length of each CMOS tube, so the classical algorithm formed by the classical logic function only expresses the operation relation of the algorithm, and does not express the dependence of the algorithm on the individuals of the CMOS tube.

The quantum logic function acts on the qubit, and the individuality of the qubit needs to be considered, such as the individuality identification, the position and the relation with surrounding qubits of the number of the qubit in the quantum chip, and the usable duration of each qubit. Therefore, the quantum algorithm formed by the quantum logic functions not only expresses the operation relation of the algorithm, but also expresses the dependence of the algorithm on quantum bit individuals.

The quantum chip can comprise quantum bits and channels for regulating the quantum bits, the quantum logic gate is realized through analog signals, and the analog signals with different combinations are applied to the quantum bits through the channels for regulating the quantum bits, so that quantum circuits with different functions are realized, and the data processing is completed. Therefore, the design of the quantum logic function acting on the quantum bit (including the design of whether the quantum bit is used or not and the design of the use efficiency of each quantum bit) is the key for improving the operation performance of the quantum computer, and special design is required, which is the uniqueness of the quantum algorithm realized based on the quantum logic function and is different from the nature and the significance of the classical algorithm realized based on the classical logic function. The above design for qubits is a technical problem that is not considered nor faced by common computing devices.

Photovoltaic power generation is a renewable energy power generation mode for directly converting light energy into electric energy by utilizing solar energy, and sunlight is mainly converted into electric energy through a solar cell panel (also called a photovoltaic cell panel and a photovoltaic panel). The photovoltaic panel is composed of a plurality of photovoltaic cells, each of which is composed of two layers of semiconductors of different materials, and when sunlight irradiates the photovoltaic cells, photons excite electrons to transition from one material to the other material, forming a current. The photovoltaic output power represents the electric energy generated by the photovoltaic panel in unit time, and is an important index for measuring the output electric energy of the photovoltaic power generation system. The actual power generation capacity and economic benefit of the photovoltaic panel are directly affected by the magnitude of the photovoltaic output power.

The photovoltaic output power prediction is to predict the photovoltaic output power of a certain period in the future according to the history and the current data, and can be divided into ultra-short-term photovoltaic power prediction, short-term photovoltaic power prediction and medium-long-term photovoltaic power prediction according to the prediction duration. Through power prediction, an operation manager of the photovoltaic power generation system can be helped to determine reasonable photovoltaic power generation capacity and distribution, the energy utilization efficiency is improved, the energy cost is reduced, and the sustainable development of renewable energy sources is promoted. Therefore, the invention provides a photovoltaic output power prediction method based on a dual-attention network and a related device, aiming at further improving the accuracy of photovoltaic output power prediction.

Referring to fig. 2, fig. 2 is a photovoltaic output power prediction method based on a dual-attention network according to an embodiment of the present invention, including the following steps:

step 201, based on photovoltaic data of a plurality of time nodes before a time node to be predicted, predicting by using a quantum long-short-time memory network to obtain a hidden state of photovoltaic output power of the time node to be predicted.

Specifically, the long-short-term memory network is a cyclic neural network, and is used for processing sequence data, and has good effects in the fields of natural language processing, voice recognition and the like. The quantum long-short time memory (Quantum Long Short Term Memory, abbreviated as QLSTM) network provides stronger computing power based on quantum superposition and quantum entanglement characteristics of quantum states by replacing a classical neural network part in a classical long-short time memory network with a variable component sub-circuit (VariationalQuantum Circuits, abbreviated as VQC), so that complex nonlinear relations between photovoltaic power data can be accurately modeled by iteratively updating parameters of a parameter-containing sub-logic gate in the variable component sub-circuit. Therefore, QLSTM networks possess higher prediction efficiency and prediction accuracy than classical long-short-term memory networks.

The hidden state is an encoded representation of past information, specifically a learning and memorization of a previously entered sequence by a network model of a cyclic structure, the hidden state being updated at each time step. Specifically, for each time step t, the long short time memory (Long Short Term Memory, LSTM) network may be configured to store the input data x corresponding to the time step _t Outputs a hidden state h _t The method comprises the steps of carrying out a first treatment on the surface of the The hidden state h _t Both the encoding of the input data of the past time step by the LSTM network and the processing of the current input data are involved. That is, the hidden state allows the LSTM network to capture long-range dependencies in the sequence by retaining long-term memory of past information while updating according to current inputs. Like classical long-short-term memory networks, QLSTM networks can be predicted from an input to be predictedAnd extracting the hidden state of the photovoltaic output power of the time node to be predicted from the photovoltaic data of a plurality of time nodes before the time node.

Step 202, processing the hidden state by using a dual-attention network to obtain a first weight and a second weight;

the first weight is the weight of the hidden state in the time dimension, and the second weight is the weight of the hidden state in the characteristic dimension.

In particular, the attention mechanism (Attention Mechanism) originates from research on human vision, which is commonly referred to as the attention mechanism in cognitive sciences, where humans can selectively focus on a portion of all information while ignoring other visible information due to bottlenecks in information processing. Before the photovoltaic output power prediction result of the time node to be predicted is obtained, the importance degree of the influence of the characteristics of the time nodes and different dimensions on the prediction result is measured, so that the learning capacity of a photovoltaic output power prediction algorithm is further improved. In the scheme provided by the embodiment of the invention, the hidden state can be processed through the dual-attention network to respectively obtain the first weight of the hidden state in the time dimension and the second weight in the characteristic dimension.

And 203, constructing a target output matrix based on the association relation among the first weight, the second weight and the hidden state.

And 204, performing full-connection decision processing on the time dimension and full-connection decision processing on the feature dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted.

Further, in the deep learning field, the fully connected layer is a densely connected structure, and each neuron is connected with all neurons of the previous layer. This means that each neuron of the fully connected layer is connected to each element in the hidden state, so that the sum of the products of each hidden state element and the corresponding weight can be calculated and converted into a probability distribution by the activation function, thereby outputting a vector matching the desired output data dimension. In this embodiment, the goal of the fully connected decision layer is to map the hidden state in the QLSTM network to the dimension of the final photovoltaic output power prediction result.

The target output matrix constructed based on the association relation among the first weight, the second weight and the hidden state not only contains the predicted information of the photovoltaic output power of the time node to be predicted by the QLSTM network, but also contains the importance degree of the characteristics of different time nodes and different dimensions on the predicted information. Then, after the dual-attention network, a full-connection layer in a time dimension and a full-connection layer in a feature dimension are added, and the full-connection decision processing in the time dimension and the full-connection decision processing in the feature dimension are sequentially performed on the target output matrix through the two full-connection layers, so that the hidden state in the QLSTM network can be mapped to the dimension of the final photovoltaic output power prediction result, and the prediction result in the dimension is output.

Therefore, in the scheme provided by the embodiment of the invention, the quantum long-short-time memory network can efficiently capture the complex nonlinear relation between the photovoltaic data by utilizing quantum superposition and quantum entanglement characteristics of quantum states and accurately model the photovoltaic data; and predicting the hidden state of the photovoltaic output power of the time node to be predicted based on the photovoltaic data of a plurality of time nodes before the time node to be predicted. The weight information in the time dimension and the characteristic dimension in the hidden state is extracted by utilizing the dual-attention network, and the change situation of a plurality of time nodes of the photovoltaic output power in the time dimension is analyzed, so that the change trend and the periodicity rule of the photovoltaic output power are fully learned; and the influence of the characteristics of different dimensions on the photovoltaic output power is learned, and a target output matrix is constructed. And the photovoltaic output power prediction result obtained through the full-connection decision processing in the time dimension and the full-connection decision processing in the feature dimension has higher accuracy.

Referring to fig. 3, as an implementation manner of the embodiment of the present invention, before the above-mentioned hidden state of the photovoltaic output power of the time node to be predicted is obtained by predicting using the quantum long short time memory network based on the photovoltaic data of the time nodes before the time node to be predicted, the above-mentioned method may further include the following steps:

Step 301, obtaining photovoltaic output power data and meteorological feature data of a plurality of time nodes before the time node to be predicted.

Specifically, in order to predict the photovoltaic output power of the time node to be predicted, photovoltaic output power data and meteorological feature data of a plurality of time nodes before the time node to be predicted can be obtained. It can be appreciated that the photovoltaic output power data tends to have periodicity, and also shows a certain trend of variation within a specific period. For example, when the weather characteristics are not greatly different, the change trend of the photovoltaic output power of each day is generally similar, and the photovoltaic output power is in an ascending trend from the sunrise in the early morning to the noon along with the gradual enhancement of the illumination intensity; from noon to sunset in the evening, the photovoltaic output power is in a descending trend until the photovoltaic output power slowly rises again at the time of the next day. Further, there is also a significant tendency for seasonal variation in photovoltaic output over a larger time dimension.

Therefore, the photovoltaic output power data of a plurality of time nodes before the time node to be predicted can also be obtained, so that the change trend of the photovoltaic output power is extracted by using the model, and the output power of the photovoltaic panel is predicted more comprehensively by combining the conditions. Specifically, the photovoltaic panel output power data can be obtained through a monitoring system or a data recording device. For example, current and voltage sensors may be installed in the photovoltaic cell stack, current and voltage values measured in real time, and photovoltaic output power values calculated.

Correspondingly, the output power of the photovoltaic power generation is also influenced by various external environment data, such as illumination intensity, temperature, geographic position, altitude and the like. For example, the most direct and important influencing factor of photovoltaic power generation is the illumination intensity, and the electric power output of a photovoltaic cell is directly related to the illumination intensity, so that the stronger the illumination, the higher the power generation. While another important factor affecting the output power is temperature: the temperature of the panel is increased to reduce the electric energy conversion efficiency of the photovoltaic cell, so that the output voltage of the cell is reduced, and the generated power is further influenced.

Therefore, in the scheme provided by the embodiment of the invention, in order to accurately predict the photovoltaic output power, at least meteorological characteristic data including illumination intensity and temperature can be extracted. Specifically, the real-time illumination intensity and temperature data can be obtained by using a weather station, a weather sensor, a solar radiation measuring instrument or other equipment. These devices may be installed at or near the photovoltaic panel for monitoring and recording environmental conditions.

Preferably, in this embodiment, the sampling frequency of the weather-characteristic data and the photovoltaic output power data should be the same. Based on the photovoltaic output power data and the meteorological feature data of the time sequence L comprising the current time node and a plurality of historical time nodes, the photovoltaic output power of the L+1 time node, namely the time node to be predicted, can be predicted; in the time sequence K, the time interval between adjacent historical time nodes is the same as the time interval from the current time node to the time node to be predicted. For example, in the ultra-short term photovoltaic power generation power prediction, if the photovoltaic output power after 1 hour in the future needs to be predicted, the time resolution may be set to 1 hour, and the length of the time sequence L may be set to 5, so as to obtain the photovoltaic output power data and the weather feature data of 5 historical time nodes in total, that is, 1 hour before, 2 hours before, 3 hours before and 4 hours before the current time. The same sampling frequency enables the photovoltaic output power prediction model to accurately extract hidden characteristic information among data, so that more accurate photovoltaic output power prediction is made.

Step 302, constructing a track matrix of the photovoltaic output power based on the photovoltaic output power data of the plurality of time nodes, and performing singular value decomposition on the track matrix.

Specifically, the track matrix refers to a matrix that can represent a time sequence evolution track by dividing time sequence data into a plurality of subsequences according to a certain time window and then stacking the subsequences together according to a time sequence, wherein each column in the track matrix represents one subsequence, and each row represents one time point in the subsequence.

Based on the obtained photovoltaic output power data of a plurality of time nodes before the time node to be predicted, a photovoltaic output power data sequence can be obtained, for example, Y _N ＝(y ₁ ,y ₂ ,...,y _N ) Wherein y is _N Photovoltaic output power data representing the nth time node can be obtained from the D fixed-length subsequences, and a track matrix X is constructed:

where L is the number of subsequences and D is the sliding window length.

Further, the trajectory matrix X may be subjected to singular value decomposition, and singular value decomposition (Singular Value Decomposition, SVD) is a matrix decomposition method. For a real or complex matrix, SVD decomposes it into the product of three matrices, i.e., x=uΣvζ, where X is an lxd matrix, U is an lxl unitary matrix, i.e., left singular matrix, Σ is an lxd diagonal matrix, i.e., singular value matrix, and vζ is a dxd unitary matrix, i.e., right singular matrix. Through SVD decomposition, the original matrix can be represented as a linear combination of a set of eigenvectors and singular values, thereby better understanding and manipulating the structure and properties of the matrix.

And 303, performing feature grouping and diagonal averaging on the submatrices obtained by singular value decomposition to obtain a power trend sequence, a power cycle sequence and a noise sequence of the plurality of time nodes.

Further, after singular value decomposition is performed on the matrix X, the following steps are performed:

X＝U∑V ^T

wherein U is E R ^L×L ，∑∈R ^L×D ，V∈R ^D×D 。

The feature grouping refers to dividing the product of the left singular matrix, the right singular matrix and the singular value matrix into a plurality of sub-matrices which are independent of the linearity of the product according to certain requirements and rules. By grouping the singular values, important feature combinations can be selectively reserved, so that the dimension of the data is reduced, and main information of the data is extracted. Wherein the trend part is a part describing a long-term trend of the time-series data in the product of the left singular matrix, the right singular matrix, and the singular value matrix, and corresponds to a low frequency region in the frequency spectrum, and represents a long-term change trend, such as a linear increase or decrease, in the time-series data. The trend portions generally correspond to smaller singular values because they correspond to low energy components in the spectrum. The period part is a component describing periodic variation in time series data in the product of the left singular matrix, the right singular matrix and the singular value matrix. It corresponds to the medium-high frequency region in the spectrum, representing periodic variations in time series data. The periodic portions generally correspond to larger singular values because they correspond to high energy components in the spectrum.

Based on the feature grouping scheme, multiple sets of submatrices may be obtained, e.g., the feature grouping may result in three sets including trend, period, and noise. Further, the diagonal averaging process is an operation for smoothing the matrix, and by averaging each element of the matrix with the adjacent elements on the diagonal thereof, noise or variation in the smoothing matrix can be reduced, and in order to prevent loss of detail, the relationship between smoothing and retaining detail needs to be weighed. The diagonal averaging process may convert the matrix into a sequential form, whereby a power trend sequence, a power cycle sequence, and a noise sequence for a plurality of time nodes may be obtained. For example, the power trend sequence may be Y _trend ＝(y _1,trend ,y _2,trend ,...,y _N,trend ) The power cycle sequence may be Y _periodic ＝(y _1,periodic ,y _2,periodic ,...,y _N,periodic )。

And step 304, carrying out normalization processing on the meteorological characteristic data, the photovoltaic output power data, the power trend sequence, the power cycle sequence and the noise sequence to obtain the photovoltaic data.

Specifically, the weather feature data, the photovoltaic output power data, the power trend sequence, the power cycle sequence and the noise sequence of the plurality of time nodes may be normalized, and in one embodiment, the element x' obtained after the normalization of any element x may be:

Wherein x is _max Is the maximum value of the elements, x _min Is the minimum value of the above elements. The data after normalization processing is the photovoltaic data of a plurality of time nodes before the time node to be predicted in the step 201.

In this embodiment, the power trend sequence, the power cycle sequence and the noise sequence of a plurality of time nodes can be obtained by performing singular value decomposition, feature grouping and diagonal averaging on the track matrix constructed by the photovoltaic output power data, so as to assist the photovoltaic output power prediction model in performing time sequence prediction. And weather characteristic data are also obtained, and the influence of the weather characteristic information on the photovoltaic output power is learned and understood, so that the prediction accuracy can be further improved.

Referring to fig. 4, as an implementation manner of the embodiment of the present invention, the above-mentioned hiding state of the photovoltaic output power of the time node to be predicted obtained by using quantum long and short time memory network prediction based on the photovoltaic data of a plurality of time nodes before the time node to be predicted may include the following steps:

and 2011, vector stitching is performed on the photovoltaic data of the time node and the hidden state corresponding to the time node according to a preset proportion for each time node before the time node to be predicted, so as to obtain a corresponding stitching vector. Wherein, the hidden state corresponding to the first time node is a preset value.

Specifically, a quantum long-short-term memory network provided by the embodiment of the invention may be shown in fig. 5, where x is as follows _t Representing the photovoltaic data of the t time node before the time node to be predicted, h _t Representing the hidden state of the t-th time node, then the photovoltaic datax _t Corresponding to the hidden state h of the t-1 time node _t-1 . Namely, the photovoltaic data of each time node corresponds to the hidden state of the last time node of the time node, and the splicing vector is the photovoltaic data x _t And hidden state h _t-1 And splicing the obtained vectors according to a preset proportion.

And 2012, inputting the spliced vector into the quantum long-short time memory network, and operating the quantum long-short time memory network to obtain the hidden state corresponding to the next time node of the time node.

In a quantum long and short-term memory network, a cell state is an internal state used for storing and transmitting information, and is a key component of the quantum long and short-term memory network, so as to solve the problem of long-term dependence and control the flow of information. The cell state is updated at each time node and is passed on to the next time node in a round robin fashion, with the definition and update of the cell state being controlled by the gating mechanism of the quantum long short duration memory network to decide which information needs to be forgotten or retained and optionally to introduce new information from the input. The definition and updating process of the cell state enables the quantum long-short-term memory network to capture and memorize long-term dependency more effectively. As shown in FIG. 5, wherein C _t Representing the cell state of the t-th time node, C _t-1 The cell status of the t-1 time node is shown.

In fig. 5, the first VQC represents the forgetting gate of the quantum long short term memory network, the second VQC represents the input gate, the third VQC represents the memory gate, and the fourth VQC represents the output gate. Then, for the forgetting gate, the input gate, the memory gate and the output gate, vector splicing can be performed on the photovoltaic data of each time node and the hidden state corresponding to the time node according to a preset proportion, so as to obtain a corresponding spliced vector, the spliced vector is used as an input of the variable component sub-circuit, and the intermediate state inside the corresponding model is obtained through calculation according to the following formula, including:

f _t ＝σ(VQC _f (w _fx x _t +w _fh h _t-1 ))

i _t ＝σ(VQC _i (w _ix x _t +w _ih h _t-1 ))

O _t ＝σ(VQC _o (w _ox x _t +w _on h _t-1 ))

wherein w is _fx For forgetting the gate VQC _f Weights, w, of photovoltaic data for the t-th time node _fh Weights for the forget gate with respect to hidden states; w (w) _ix For input gate VQC _i Weights, w, of photovoltaic data for the t-th time node _ih Weights for the input gate with respect to hidden states; w (w) _ux For memory gate VQC _u Weights, w, of photovoltaic data for the t-th time node _uh Weights for the memory gate with respect to hidden states; w (w) _ox For the output gate VQC _o Weights, w, of photovoltaic data for the t-th time node _oh The weight of the gate with respect to the hidden state is output. Sigma represents the activation function sigmoid, tanh represents the hyperbolic tangent activation function tanh, and the quantum long and short time memory network also comprises an addition operationSum product operation->

It can be understood that the quantum long-short-term memory network is a network model of cyclic iteration, and the iterative process corresponding to each time node is as follows: the QLSTM network constructs a spliced vector based on the photovoltaic data of each time node and the hidden state of the previous time node, and inputs the spliced vector into a forgetting gate, an input gate, a memory gate and an output gate respectively; processing the VQC obtained through training, and obtaining an intermediate state through activating a function; and information flow of forgetting, inputting, memorizing and outputting processes is carried out under the control of the cell state of the previous time node, so as to obtain the hidden state of the time node, namely the hidden state corresponding to the next time node of the time node (used for constructing the splicing vector of the next time node), and the cell state of the time node (used for controlling the information flow corresponding to the next time node). Since the first time node has no previous time node, the hidden state corresponding to the first time node can be set to a preset value according to experience.

And step 2013, determining the hidden state of the photovoltaic output power of the time node to be predicted based on the hidden states corresponding to all the time nodes.

Based on the scheme, after the photovoltaic data of each time node is input into the vector long-short-time memory network, the hidden state of the time node can be obtained, and then the hidden state of each time node can be obtained through repeated loop iteration, so that the hidden state of the photovoltaic output power of the time node to be predicted can be determined based on the hidden state corresponding to each time node. In one embodiment, the hidden state h of each time node except the first time node can be determined _t The hidden characteristic included in the hidden state is used for representing, and as the hidden characteristic of each time node can comprise a plurality of dimensions, the dimensions can be ordered according to a preset sequence, so that a plurality of feature vectors corresponding to a plurality of hidden features with the same dimension sequence are obtained; the feature vectors are sequentially used as column vectors of a matrix according to a time sequence, so that a hidden feature matrix can be constructed, and the hidden feature matrix can be used as a hidden state of the photovoltaic output power of the time node to be predicted.

As an implementation manner of the embodiment of the present invention, the quantum long-short time memory network may include a plurality of variable component sub-circuits, and a specific structure of each variable component sub-circuit may be as shown in fig. 6, where each variable component sub-circuit includes a coding circuit and a variable component circuit connected in sequence, the coding circuit includes a first single quantum logic gate combination acting on each qubit, and the variable component circuit includes a second single quantum logic gate combination acting on each qubit and a multiple quantum logic gate acting on a plurality of qubits, where:

Specifically, as shown in fig. 6, the variable component sub-circuit may include the same number of qubits as the number of dimensions of the hidden feature, both being 4, i.e., the variable component sub-circuit includes the qubits q ₀ 、q ₁ 、q ₂ And q ₃ And the hidden characteristic is used for encoding hidden characteristics of each dimension respectively, so that the spliced vectors can be subjected to angle encoding through the cascade H gate, the first RY gate and the first RZ gate which act on each quantum bit, and the quantum state corresponding to each spliced vector is obtained.

The rotation parameters of the second RZ gate, the second RY gate, the third RZ gate and the CRX gate are trainable parameters, and the rotation parameters need to be adjusted in advance through training to optimize each VQC. After quantum state coding is completed, the quantum state obtained by the coding can be entangled and evolved through a second RZ gate, a second RY gate, a third RZ gate and a CRX gate which are cascaded, and the complex nonlinear relation between the photovoltaic data information is captured by utilizing quantum superposition and quantum entanglement characteristics of the quantum state. The CRX gate is a parameter-containing two-bit quantum gate, and the two-bit quantum gate with parameters is used for exploring trainable entanglement, has self-adaptive characteristics and more flexible entanglement capacity, can be better used for QNN (Quantum Neural Networks, quantum neural network) algorithms of different tasks, and the VQC with the two parameterized quantum bit gates can learn mapping functions in a larger search space, so that photovoltaic output power can be predicted more accurately.

Based on the algorithm flow and the specific quantum circuit structure, the QLSTM may output hidden states of tensors in the form of [ batch, seq_len, feature_size [ hidden_size ] ], where batch represents the number of samples processed in one loop iteration; seq_len represents the length of the input photovoltaic data, i.e. the number of the above-mentioned plurality of time nodes; feature_size or hidden_size represents the feature dimension of the hidden state corresponding to each time node. Further, the hidden state needs to be input into the dual-attention network for further processing.

In this embodiment, the hidden states of the photovoltaic output power of the time node to be predicted can be obtained through the QLSTM network by using the hidden states of all the time nodes, so that the change trend and the periodicity rule of the photovoltaic output power are fully learned, and the influence of the characteristics of different dimensions on the photovoltaic output power is understood. The quantum neural network based on the parameter-containing double-bit quantum gate is designed, the expression capacity of a photovoltaic processing power prediction model is improved, and meanwhile, the generalization capacity of the model is improved.

As one implementation of the embodiment of the present invention, the dual attention network may include a time attention network and a feature attention network;

The processing the hidden state by using the dual-attention network to obtain the first weight and the second weight may include:

and extracting time weight components on each time node in the hidden state by using the time attention network to obtain a first weight.

Specifically, when the QLSTM network outputs the hidden state of the photovoltaic output power of the time node to be predicted; for example, when the hidden state is a hidden feature matrix including tensors of seq_len and feature_size, an input matrix I of the dual-attention network may be constructed:

I∈h ^s×f

wherein s represents the number of the plurality of time nodes before the time node to be predicted, f represents the dimension number of the hidden feature corresponding to each time node, namely, the input matrix I is a hidden feature matrix of f rows and s columns, and each element in the matrix is a state component of the hidden state of the photovoltaic output power of the time node to be predicted.

In this embodiment, the hidden state may be a hidden feature matrix constructed based on hidden features of a plurality of time nodes, where each column vector of the hidden feature matrix corresponds to a hidden feature of a time node, and each row vector corresponds to a hidden feature of a feature dimension;

The extracting, by using the time attention network, the time weight components on each time node in the hidden state to obtain the first weight may include the following steps:

and carrying out average pooling on elements in each column vector in the hidden characteristic matrix to obtain a one-dimensional time array corresponding to the hidden characteristic matrix.

And carrying out convolution processing on the one-dimensional time array to obtain time weight components corresponding to each time node, and constructing the first weight based on the time weight components.

Further, for each column vector (i.e. hidden feature corresponding to each time node) in the input matrix I, the elements therein may be averaged and pooled, i.e. matrix element { h } is calculated ₁₁ 、h ₂₁ 、h ₃₁ …h _f1 Mean value of z ₁ The method comprises the steps of carrying out a first treatment on the surface of the Calculating matrix element { h ₁₂ 、h ₂₂ 、h ₃₂ …h _f2 Mean value of z ₂ The method comprises the steps of carrying out a first treatment on the surface of the …; until the matrix element { h } is calculated _1s 、h _2s 、h _3s …h _fs Mean value of z _f . Thereby constructing a one-dimensional time array [ z ] corresponding to the input matrix I ₁ ,z ₂ ,…,z _f ]And then carrying out convolution processing procedures of one-dimensional convolution, a Relu activation function, one-dimensional convolution and a Sigmoid activation function on the one-dimensional time array in sequence to obtain a weight vector. Each component of the weight vector is a time weight component corresponding to each time node, and represents the influence degree of the corresponding time node on the photovoltaic output power prediction result; the larger the time weight component, the greater the contribution of the time node in predicting the photovoltaic output power, so that the above-described first weight can be constructed on a per time weight component basis.

Correspondingly, the extracting the feature weight components on each feature dimension in the hidden state by using the feature attention network to obtain the second weight may include the following steps:

and carrying out average pooling on elements in each row of vectors in the hidden feature matrix to obtain a one-dimensional feature array corresponding to the hidden feature matrix.

Similarly, for each row of vectors (i.e. hidden features corresponding to each feature dimension) in the input matrix|, the elements in the vectors can be averaged and pooled, i.e. the average value of the elements of the corresponding calculation matrix can be constructed, and a one-dimensional feature array [ x ] corresponding to the input matrix I can be constructed ₁ ,x ₂ ,…,x _s ]And then carrying out convolution processing on the one-dimensional feature array to obtain another weight vector. Each component of the weight vector is a characteristic weight component corresponding to each characteristic dimension, and the influence degree of the corresponding characteristic dimension on the photovoltaic output power prediction result is represented; the larger the feature weight components, the larger the contribution of the feature dimension in predicting the photovoltaic output power, so that the above-mentioned second weight can be constructed based on each feature weight component.

As shown in fig. 7, as an implementation manner of the embodiment of the present invention, the above-mentioned construction of the target output matrix based on the association relationship between the first weight, the second weight and the hidden state may include the following steps:

step 701, constructing a first output matrix based on the time weight component and the state components in the hidden state.

Specifically, in this embodiment, the constructing the first output matrix based on the time weight component and the state component in the hidden state may include:

and calculating the product of each element in the hidden characteristic matrix and the time weight component of the corresponding time node to be used as the element in the first output matrix to obtain the first output matrix.

Specifically, the hidden feature matrix is constructed based on hidden features of a plurality of time nodes, wherein each column vector corresponds to a hidden feature of a time node, and each row vector corresponds to a hidden feature of a feature dimension, so that one element in the hidden feature matrix can represent a component of the hidden feature of a specific time node in a specified feature dimension, namely, one state component of the hidden state of the photovoltaic output power of the time node to be predicted; in the first weight, the time weight component of the corresponding time node indicates the influence degree of the time node on the photovoltaic output power prediction result; the product of the two can accurately measure the contribution of each state component to the photovoltaic output power prediction result in the time dimension, and the constructed first output matrix can comprehensively and accurately embody the influence of different time nodes on the photovoltaic output power prediction result.

In one embodiment, a plurality of first weights may be copied and a time weight matrix is obtained based on the plurality of first weights, the time weight matrix has the same scale as the input matrix I, and the hadamard product of the time weight matrix and the input matrix I is calculated to complete the product calculation of each element in the hidden feature matrix and the time weight component of the corresponding time node, thereby obtaining a first output matrix U _s 。

Step 702, constructing a second output matrix based on the feature weight components and the state components in the hidden state.

Correspondingly, the constructing a second output matrix based on the feature weight component and the state component in the hidden state may include:

Similarly, in the second weight, the characteristic weight component of the corresponding characteristic dimension indicates the influence degree of the characteristic dimension on the photovoltaic output power prediction result, so that the product of the characteristic dimension and each element in the hidden characteristic matrix can accurately measure the contribution of each state component on the characteristic dimension of the photovoltaic output power prediction result, and the constructed second output matrix can comprehensively and accurately reflect the influence of different characteristic dimensions on the photovoltaic output power prediction result.

Correspondingly, in one embodiment, a plurality of second weights may be duplicated, and a feature weight matrix is obtained based on the plurality of second weights, the scale of the feature weight matrix is the same as that of the input matrix I, and the hadard product of the feature weight matrix and the input matrix I is calculated, so as to complete the product calculation of each element in the hidden feature matrix and the feature weight component of the corresponding feature dimension, thereby obtaining a second output matrix U _f 。

Step 703, calculating the sum of the first output matrix and the second output matrix as the target output matrix.

Further, the first output matrix U may be calculated _s And a second output matrix U _f The sum, as the target output matrix, is, for example, QLSTM output shape such as [ batch, seq_len, feature_size ]]When the hidden state of the tensor of (a), i.e. the hidden state comprises s×f feature components, an input matrix I of the dual-attention network can be constructed, then a first output matrix U determined based on the temporal weight component, the feature weight component and the feature component _s And a second output matrix U _f Also of scale f x s, the last first output matrix U _s And a second output momentArray U _f Target output matrix U of sum _sf Still hold [ batch, seq_len, feature_size ]]Thereby guaranteeing the integrity of the data information in the hidden state.

Because the target output matrix is constructed based on the association relation among the first weight, the second weight and the hidden state, the target output matrix not only contains the predicted information of the photovoltaic output power of the time node to be predicted by the QLSTM network, but also contains the importance degree of the characteristics of different time nodes and different dimensions on the predicted information. And the full connection decision processing in the time dimension and the full connection decision processing in the characteristic dimension are sequentially carried out on the target output matrix through the full connection layer in the time dimension and the full connection layer in the characteristic dimension, so that the hidden state in the QLSTM network can be mapped to the dimension of the final photovoltaic output power prediction result, and finally the photovoltaic output power prediction result in the dimension is output through the inverse normalization processing of the mapping result, thereby further improving the accuracy of photovoltaic output power prediction.

In one embodiment, the photovoltaic output power of the time node to be predicted can be respectively predicted through a plurality of circulating network models, such as a classical LSTM network model, a QLSTM network model, and a model based on a dual-attention network and a QLSTM network provided by the scheme; and measuring the photovoltaic output power of the time node to be predicted to obtain a true value. Therefore, a predicted result of each network model and a line graph of the real value of the photovoltaic output power changing along with time can be drawn, wherein the abscissa is time information, and the ordinate is photovoltaic output power information.

As shown in fig. 8, the prediction results of the above-mentioned various photovoltaic output power prediction modes may be shown, where the broken line (1) is a predicted value based on a dual-attention network and a QLSTM network according to the technical solution of the present invention, the broken line (2) is a predicted value of the QLSTM network, the broken line (3) is a predicted value of a classical LSTM network, and the broken line (4) is a measured actual value of the photovoltaic output power. It can be seen that the photovoltaic output power fluctuates by a large extent during the time period included in this line graph, and that the weather may be cloudy or even overcast at that time. Compared with classical LSTM networks and QLSTM networks, the predicted value of the technical scheme of the invention is always closer to the true value of the photovoltaic output power, so that the predicted model provided by the technical scheme of the invention has higher accuracy of predicting the photovoltaic output power even in an environment with larger change of climate conditions.

Referring to fig. 9, fig. 9 is a photovoltaic output power prediction apparatus based on a dual-attention network according to an embodiment of the present invention, where the apparatus includes:

the prediction module 901 is configured to predict, based on photovoltaic data of a plurality of time nodes before a time node to be predicted, a hidden state of photovoltaic output power of the time node to be predicted by using a quantum long-short time memory network.

A weight module 902, configured to process the hidden state by using a dual-attention network to obtain a first weight and a second weight; the first weight is the weight of the hidden state in the time dimension, and the second weight is the weight of the hidden state in the characteristic dimension.

The output module 903 is configured to construct a target output matrix based on the association relationship between the first weight, the second weight, and the hidden state.

And the decision module 904 is used for performing full-connection decision processing on the time dimension and full-connection decision processing on the feature dimension on the target output matrix to obtain a photovoltaic output power prediction result of the time node to be predicted.

With respect to specific functions and effects achieved by the photovoltaic output power prediction device based on the dual-attention network, reference may be made to other embodiments of the present specification for comparison and explanation, and no further description is given here. The individual modules in the dual-attention network based photovoltaic output power prediction device may be implemented in whole or in part by software, hardware, and combinations thereof. The modules can be embedded in hardware or independent of a processor in the computer equipment, and can also be stored in a memory in the computer equipment in a software mode, so that the processor can call and execute the operations corresponding to the modules.

Please refer to fig. 10. The embodiment of the present specification further provides a computer device, including a memory and a processor, where the memory stores a computer program, and the processor executes the computer program to implement the photovoltaic output power prediction method based on the dual-attention network in any of the above embodiments. Referring to fig. 10, the computer device may be a classical computer. The computer device may also be a quantum computer.

The present description also provides a computer-readable storage medium having stored thereon a computer program which, when executed by a computer, causes the computer to perform the dual-attention network-based photovoltaic output power prediction method of any of the above embodiments.

Embodiments of the present disclosure also provide a computer program product comprising instructions that, when executed by a computer, cause the computer to perform the dual-attention network-based photovoltaic output power prediction method of any of the above embodiments.

It will be appreciated that the specific examples in this specification are intended only to assist those skilled in the art in better understanding the embodiments of the present specification and are not intended to limit the scope of the invention.

It should be understood that, in various embodiments of the present disclosure, the sequence number of each process does not mean that the execution sequence of each process should be determined by its functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

It will be appreciated that the various embodiments described in this specification may be implemented either alone or in combination, and are not limited in this regard.

Unless defined otherwise, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this specification belongs. The terminology used in the description is for the purpose of describing particular embodiments only and is not intended to limit the scope of the description. The term "and/or" as used in this specification includes any and all combinations of one or more of the associated listed items. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

It will be appreciated that the processor of the embodiments of the present description may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method embodiments may be implemented by integrated logic circuits of hardware in a processor or instructions in software form. The processor may be a general purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), an off-the-shelf programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components. The methods, steps and logic blocks disclosed in the embodiments of the present specification may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of a method disclosed in connection with the embodiments of the present specification may be embodied directly in hardware, in a decoded processor, or in a combination of hardware and software modules in a decoded processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in a memory, and the processor reads the information in the memory and, in combination with its hardware, performs the steps of the above method.

It will be appreciated that the memory in the embodiments of this specification may be either volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile memory may be a read-only memory (ROM), a Programmable ROM (PROM), an Erasable Programmable ROM (EPROM), an Electrically Erasable Programmable ROM (EEPROM), or a flash memory, among others. The volatile memory may be Random Access Memory (RAM). It should be noted that the memory of the systems and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present specification.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described system, apparatus and unit may refer to corresponding procedures in the foregoing method embodiments, and are not repeated herein.

In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods may be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiment.

In addition, each functional unit in each embodiment of the present specification may be integrated into one processing unit, each unit may exist alone physically, or two or more units may be integrated into one unit.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present specification may be essentially or portions contributing to the prior art or portions of the technical solutions may be embodied in the form of a software product stored in a storage medium, including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the embodiments of the present specification. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a read-only memory (ROM), a random-access memory (RAM), a magnetic disk, or an optical disk, etc.

The foregoing is merely specific embodiments of the present disclosure, but the scope of the present disclosure is not limited thereto, and any person skilled in the art can easily think about changes or substitutions within the technical scope disclosed in the present disclosure, and should be covered by the scope of the present disclosure. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A photovoltaic output power prediction method based on a dual-attention network, the method comprising:

2. The method of claim 1, wherein the dual attention network comprises a temporal attention network and a feature attention network;

3. The method of claim 2, wherein the constructing a target output matrix based on the association between the first weight, the second weight, and the hidden state comprises:

constructing a second output matrix based on the feature weight components and the state components in the hidden state;

and calculating the sum of the first output matrix and the second output matrix as the target output matrix.

4. The method of claim 2, wherein the hidden state is a hidden feature matrix constructed based on hidden features of a plurality of time nodes, each column vector of the hidden feature matrix corresponding to a hidden feature of a time node, each row vector corresponding to a hidden feature of a feature dimension;

5. The method of claim 3, wherein the hidden state is a hidden feature matrix constructed based on hidden features of a plurality of time nodes, each column vector of the hidden feature matrix corresponding to a hidden feature of a time node, each row vector corresponding to a hidden feature of a feature dimension;

6. The method of any of claims 1-5, wherein prior to predicting, using a quantum long-short-term memory network, a hidden state of the photovoltaic output power of the time node to be predicted based on the photovoltaic data of a plurality of time nodes preceding the time node to be predicted, the method further comprises:

7. The method according to any one of claims 1 to 5, wherein predicting, based on the photovoltaic data of a plurality of time nodes before the time node to be predicted, a hidden state of the photovoltaic output power of the time node to be predicted using a quantum long short time memory network includes:

8. The method of claim 7, wherein the quantum long and short time memory network comprises a plurality of variable component sub-lines, each comprising a coding line and a variable component line connected in sequence, the coding line comprising a first single quantum logic gate combination acting on each qubit, the variable component line comprising a second single quantum logic gate combination acting on each qubit and a multiple quantum logic gate acting on a plurality of qubits, wherein:

9. A dual-attention network-based photovoltaic output power prediction device, the device comprising:

10. A storage medium having a computer program stored therein, wherein the computer program is arranged to perform the method of any of claims 1 to 8 when run.

11. An electronic device comprising a memory and a processor, characterized in that the memory has stored therein a computer program, the processor being arranged to run the computer program to perform the method of any of the claims 1 to 8.