High-precision electricity price prediction method based on multi-source data fusion algorithm

Abstract

The invention relates to the technical field of electric power system intelligent prediction and energy management, and discloses a high-precision electricity price prediction method based on a multi-source data fusion algorithm, and the method comprises the steps: collecting and normalizing the multi-source time sequence data of an electric power market and an energy storage system; dividing into three characteristic cabins of supply, demand and energy storage; feature subset selection is carried out based on a two-stage structure particle swarm algorithm; and finally predicting by adopting a double-layer electricity price prediction model. Compared with the prior art that feature physical semantic division and structure compression are not considered in a feature selection stage, and particularly in a multi-source high-dimensional feature scene, the technical problem that a prediction model is easy to have overfitting or weak generalization ability is easily caused, the method has the advantages that the three types of feature cabins of supply, demand and energy storage and two-stage particle structure coding are constructed, so that the prediction efficiency is improved; and in combination with single-layer and double-layer prediction model decoupling deployment, controllable compression and high-stability prediction in a high-dimensional feature space are realized, and the accuracy and generalization ability of electricity price prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent prediction of power system and energy management, and particularly relates to a high-precision electricity price prediction method based on a multi-source data fusion algorithm. BACKGROUND

[0002] At present, as a core link of energy scheduling and energy storage control, electricity price prediction widely relies on machine learning models driven by historical electricity price, load, weather and other time series data. However, most of the existing methods use unified feature pools or black box neural network structures, and fail to combine the physical semantics of various feature data in the power system for structured processing, resulting in the deficiencies of the model, such as serious feature redundancy, poor generalization ability and large fluctuation of prediction accuracy. For example, when processing multi-source heterogeneous data, traditional methods often directly splice all features for input without explicit selection and compartment-level division, which leads the model to fall into local optimum or overfitting. The existing technology cannot fully meet the prediction needs of "multi-source data fusion + feature compression control + depth modeling stability", especially in the energy storage scheduling scene with high feature dimension and uneven data quality, the performance of the prediction system is difficult to guarantee. Therefore, there is an urgent need for a high-precision electricity price prediction method that can still realize feature structured selection, redundancy control and stable modeling under the condition of multi-source high-dimensional feature fusion, so as to improve the accuracy, robustness and deployability of the prediction system. SUMMARY

[0003] In view of the above technical deficiencies, the purpose of the present application is to provide a high-precision electricity price prediction method based on a multi-source data fusion algorithm, which aims to solve the technical problems that the existing technology does not consider the feature physical semantic division and structure compression in the feature selection stage, especially in the multi-source high-dimensional feature scene, which easily leads to overfitting or weak generalization ability of the prediction model.

[0004] To solve the above technical problems, the present application adopts the following technical scheme: the present application provides a high-precision electricity price prediction method based on a multi-source data fusion algorithm,

[0005] The high-precision electricity price prediction method based on the multi-source data fusion algorithm comprises:

[0006] Step S10: collecting and preprocessing multi-source time series data associated with the electricity market and the energy storage system, dividing the multi-source time series data into three types of feature cabin data according to the physical source of the multi-source time series data, the three types of feature cabin data including supply feature cabin data G, demand feature cabin data D and energy storage feature cabin data S; normalizing the three types of feature cabin data to the interval [0, 1] and constructing a supply feature matrix X G , a demand feature matrix X D and an energy storage feature matrix X S ;

[0007] Step S20: introducing a binary and real value hybrid particle swarm optimization updating mechanism, based on the supply feature matrix X G , the demand feature matrix X D and the energy storage feature matrix X S Each particle in the particle swarm is coded as a two-level structure, and the feature subset of particle i is set according to the two-level structure and the two-level structure particle coding result is output.

[0008] Step S30: according to the feature subset , the corresponding feature matrix is extracted and input into the preset single-layer electricity price prediction model for training, and the mean absolute percentage error is calculated to construct the fitness function

[0009] Step S40: based on the fitness function , the two-level structure coding of each particle is iteratively updated, and the final optimized particle feature subset S * and the optimized weight set W * are output.

[0010] Step S50: according to the optimized particle feature subset S * and the optimized weight set W * , the weighted input matrix X final is constructed final , the weighted input matrix X G is input into the preset double-layer electricity price prediction model, and the final predicted electricity price is output.

[0011] Preferably, in step S10, the supply feature cabin data G includes wind power output data, photovoltaic output data and thermal power planned output data; the demand feature cabin data D includes total load data, temperature data and workday identification data; and the energy storage feature cabin data S includes energy storage state of charge data and charge and discharge power data.

[0012] Preferably, in step S20, based on the supply feature matrix X D , the demand feature matrix X S and the energy storage feature matrix X i , each particle in the particle swarm is coded as a two-level structure, and the two-level structure coding of particle i is wherein x i is the two-level structure coding of particle i, is the indicator value of whether particle i enables the preset supply cabin, demand cabin or energy storage cabin; is the selection bit of particle i for the features in each cabin; and W 3 is the continuous feature weight of particle i in each cabin; || represents the vector splicing operation; wherein, {0,1} GRepresents a vector of length 3, each bit has a value of 0 or 1, b G To supply the outer layer switch, b D is the outer switch of the demand cabin, b S For the energy storage cabin outer switch, only when a cabin outer switch b ( ·)=1, the cabin feature bit takes effect.

[0013] Preferably, in step S40, based on the fitness function Iteratively update the two-level structure encoding of each particle and output the final optimal particle feature subset S * and weight set W * The steps specifically include: updating the outer cabin selection position and the inner feature selection position in the two-level structure using Sigmoid combined with random sampling method; updating the continuous feature weights in the two-level structure using speed and position update formulas; recalculating the prediction error after each round of update and updating the individual optimal and global optimal particle positions, iterating until the fitness function converges or reaches the preset maximum number of iterations, and outputting the final optimal particle feature subset S * and weight set W * .

[0014] Preferably, in step S30, the preset single-layer electricity price prediction model is constructed based on the long short-term memory network method.

[0015] Preferably, in step S50, the double-layer electricity price prediction model adopts a comprehensive loss function during the training process. for: Where T is the total length of the training sample; τ is the prediction time step; P(t+τ) is the electricity price at the corresponding time in the preset verification set at time t+τ; is the electricity price predicted by the model at time t+τ; β is the variance regularization penalty coefficient, which is used to suppress prediction fluctuations; Var is the variance of the predicted electricity price, which is used to control the prediction stability.

[0016] Preferably, in step S50, the structure of the two-layer electricity price prediction model includes an input layer, a first long short-term memory network layer, a second long short-term memory network layer, a fully connected layer and an output layer, wherein the first long short-term memory network layer is used to extract short-term dynamic dependencies from the weighted feature matrix; the second long short-term memory network layer is used to learn long-term trend information in the electricity price time series;

[0017] Among them, the second-layer long short-term memory network layer is consistent with the single-layer electricity price prediction model structure trained in step S30, and the model parameters trained in step S30 are used as the initial parameters of the second-layer long short-term memory network to realize the migration of feature subsets in the lightweight training stage and the deep modeling stage.

[0018] The present invention also provides a high-precision electricity price prediction method system based on a multi-source data fusion algorithm, comprising:

[0019] The feature preprocessing module is used to collect and preprocess the multi-source time series data associated with the power market and the energy storage system. The multi-source time series data is divided into three types of feature cabin data according to their physical sources. The three types of feature cabins include supply feature cabin data G, demand feature cabin data D, and energy storage feature cabin data S. The three types of feature cabin data are normalized to the interval [0, 1] and the supply feature matrix X is constructed. G , demand feature matrix X D and energy storage characteristic matrix X S ;

[0020] The two-level structure particle encoding module is used to introduce a binary and real-valued hybrid particle swarm update mechanism based on the supply feature matrix X G , demand feature matrix X D and energy storage characteristic matrix X S Each particle in the particle swarm is encoded into a two-level structure, and the feature subset of particle i is set according to the constructed two-level structure. And output the two-level structure particle encoding results;

[0021] Single-layer prediction fitness evaluation module is used to evaluate the performance of a feature subset Extract the corresponding feature matrix And input it into the preset single-layer electricity price prediction model for training, and calculate the mean absolute percentage error Constructing a fitness function

[0022] Particle swarm iterative optimization module is used to optimize the Iteratively update the two-level structure encoding of each particle and output the final optimized particle feature subset S * and optimize the weight set W * ;

[0023] The two-layer prediction deployment module is used to optimize the particle feature subset S * and optimize the weight set W * Construct the weighted input matrix X final , the weighted input matrix X final Input into the preset two-layer electricity price prediction model, and the model outputs the final predicted electricity price.

[0024] The present invention also provides a computer program product, including a high-precision electricity price prediction method program based on a multi-source data fusion algorithm. When the high-precision electricity price prediction method program based on a multi-source data fusion algorithm is executed by a processor, the high-precision electricity price prediction method based on a multi-source data fusion algorithm is implemented.

[0025] The beneficial effects of the present application are that by introducing three types of physical feature cabin division mechanism and two-level structure particle coding method, the structured selection and compression control of multi-source high-dimensional time sequence characteristics are realized, the input variable dimension is effectively reduced, the explainability of feature selection is improved, and the redundancy interference problem existing in the traditional "unified splicing" mode is avoided.

[0026] A collaborative prediction architecture composed of a single-layer lightweight prediction model and a double-layer deep modeling model is constructed, which can quickly evaluate the subset performance in the feature optimization stage, realize deep fitting of the time-dependent structure in the final deployment stage, significantly improve the prediction accuracy and stability of the model, and is suitable for the energy storage scheduling scene with high requirements for generalization ability and robustness in electricity price prediction. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0028] Figure 1 The flowchart of the first embodiment of the present application is shown in the first embodiment of the present application.

[0029] Figure 2 The device schematic diagram of the present application is shown in the first embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0031] Embodiment one: as shown in the first embodiment of the present application, the flowchart of the first embodiment of the present application is shown in the first embodiment of the present application. Figure 1

[0032] In the first embodiment, the high-precision electricity price prediction method based on multi-source data fusion algorithm comprises:

[0033] ​Step S10: Collect and pre-process the multi-source time-series data associated with the power market and energy storage system, divide it into three types of feature cabin data according to the physical source of the multi-source time-series data, the three types of feature cabin data including supply feature cabin data G, demand feature cabin data D and energy storage feature cabin data S; normalize the three types of feature cabin data to the interval [0, 1] and construct the supply feature matrix X G , the demand feature matrix X D and the energy storage feature matrix X S ;

[0034] It should be noted that, on the basis of the traditional multi-source electricity price prediction method, the "feature cabin" concept is introduced in this step, that is, logical division is carried out according to the physical system source of the data, which is different from the way of uniformly flattening all variables in the prior art. This division not only retains the semantic structure of the data in the system (such as the difference between the supply and demand sides, the response characteristics of the energy storage side), but also provides a structured coding basis for the subsequent particle swarm algorithm, significantly reducing the feature dimension redundancy and computational complexity. The normalization step uses the minimum-maximum standardization formula based on the sliding time window, which is suitable for feature variables such as SOC and electricity price in the energy storage system that have boundary restrictions and periodic jump characteristics, and helps to improve the numerical stability of the prediction model.

[0035] Step S20: Introduce a binary and real-valued hybrid particle swarm update mechanism, and based on the supply feature matrix X G , the demand feature matrix X D and the energy storage feature matrix X S , the two-level structure of each particle in the particle swarm is coded, and the feature subset of particle i is set according to the two-level structure, and the two-level structure particle coding result is output;

[0036] It should be noted that, in step S20, based on the supply feature matrix X G , the demand feature matrix X D and the energy storage feature matrix X S , the two-level structure of each particle in the particle swarm is coded, and the two-level structure coding of particle i is wherein x i is the two-level structure coding of particle i, is the indicator value of whether particle i enables the preset supply cabin, demand cabin or energy storage cabin; is the selection bit of particle i to the features in each cabin; W i is the weight of particle i to the continuous features in each cabin; || represents the vector splicing operation; wherein {0,1} 3 represents a vector of length 3, each bit taking the value 0 or 1, b G is the outer switch of the supply cabin, b D is the outer switch of the demand cabin, bS For the energy storage out-of-cabin switch, only when the out-of-cabin switch b (·) =1, its cabin feature bit is effective. This step breaks the single-layer bit string structure of "global plane feature space selection" in the traditional particle swarm optimization algorithm by introducing a two-level structure particle coding mechanism, so that the feature selection process has "cabin-level control" and "in-cabin compression" capabilities. This structure makes full use of the supply, demand and energy storage feature matrices constructed in step S10, realizes the interpretable selection logic facing the physical cabin semantics, and has stronger structural constraint ability and redundant feature compression ability.

[0037] It can be understood that each particle under this structure not only represents a candidate feature subset, but also contains cabin-level structure control and feature weight vector, laying a foundation for subsequent construction of weighted input matrix and construction of fitness function. Compared with the traditional particle swarm optimization algorithm which only does simple gating on input variables, this structure introduces physical structure prior in the power system scenario, has stronger semantic interpretation ability and model robustness, and is especially suitable for high-dimensional multi-source, multi-associated time series feature selection problems.

[0038] For example, X G , X D , X S contains 16, 18, 14 features respectively, if the outer coding of particle i is [1, 1, 0], it means that only the supply cabin and demand cabin are enabled in this round of feature selection, and the energy storage cabin is disabled; if the inner coding of the particle selects the 1st, 3rd and 5th features of the supply cabin and the 2nd, 4th and 6th features of the demand cabin, the input dimension corresponding to the particle is 6, which is much smaller than the full-splicing feature dimension of 48. In a certain day-ahead market simulation, compared with the traditional PSO feature selection method, the training time of this structure is reduced by 32%, and the validation error is reduced by about 3.8%, which verifies its dual effects in feature space compression and prediction accuracy improvement.

[0039] Step S30: According to the feature subset extract the corresponding feature matrix and input it into the preset single-layer electricity price prediction model for training, and calculate the mean absolute percentage error to construct the fitness function

[0040] It should be noted that, based on the particle encoding result output in step S20, the corresponding feature subset is extracted and a feature matrix is ​​constructed, and then the matrix is ​​used as input and passed into the set lightweight single-layer electricity price prediction model for training and verification to obtain the prediction performance index corresponding to the particle. By calculating the mean absolute percentage error of the model on the verification set as the loss index, a fitness function for the global search of the particle swarm is further constructed, thereby achieving the quality judgment of the feature subset. The single-layer prediction model used in this step is preferably a single-layer LSTM neural network, which has good temporal memory ability and parameter compression characteristics, and is suitable for fast training and performance evaluation when a large number of particles are input. The model structure is shallow, has a small number of parameters, and can converge quickly, which is convenient for repeated use in the iterative evaluation of each particle. As a "candidate evaluator" in the entire search framework, it does not introduce a computational bottleneck. The fitness function not only includes the prediction error term, but also introduces a sparsity regularization term according to the feature compression ratio of the particle, and constructs a fitness function in the following form: Among them, F i is the fitness value of particle i; E i The error of the corresponding prediction model on the validation set; λ<0, λ is the feature redundancy penalty weight; b inner ∈{0,1} is the feature selection bit in the particle inner layer bit string; N all is the original total feature dimension; this function can compress redundant dimensions while ensuring prediction accuracy and improve the generalization ability of the final model.

[0041] Step S40: Based on the fitness function Iteratively update the two-level structure encoding of each particle and output the final optimized particle feature subset S * and optimize the weight set W * ;

[0042] It should be noted that, in step S40, based on the fitness function Iteratively update the two-level structure encoding of each particle and output the final optimal particle feature subset S * and weight set W * The steps specifically include: updating the outer cabin selection position and the inner feature selection position in the two-level structure using Sigmoid combined with random sampling method; updating the continuous feature weights in the two-level structure using speed and position update formulas; recalculating the prediction error after each round of update and updating the individual optimal and global optimal particle positions, iterating until the fitness function converges or reaches the preset maximum number of iterations, and outputting the final optimal particle feature subset S * and weight set W * .

[0043] It can be understood that the binary and real value hybrid particle updating mechanism enables the particle to realize "bit-level search" on whether to enable a certain cabin and the characteristics therein, and "real value fine-tuning" on the weight intensity of the enabled characteristics in the structure space, thereby synchronously completing the joint modeling of the structure optimization of the characteristic subset and the importance ranking of the characteristics in the global search process. This mechanism avoids the limitation of only having binary strings and lacking intensity information in traditional PSO feature selection, and significantly improves the convergence ability of the final model to the minimum redundant feature combination

[0044] For example, assuming that the current particle swarm size is 30, the maximum number of iterations is 50, and the initial particles are randomly activated and initialized with feature bits. At the 27th round, the particle swarm converges to the global optimal particle x * , and the encoding result is: outer bit: [b G ,b D ,b S ] = [1, 1, 0], inner selection bit: a total of 14 feature dimensions are enabled; weight set W * = [0.92, 0.76, 0.31,...]. Compared with the full feature (48-dimensional) model, the test error is reduced by about 27%, and because of its stable structure, the variance fluctuation on different validation sets is less than 1.2%, proving that it has good prediction accuracy improvement ability and stable generalization ability. The optimized feature subset S * and the weight set W * will be used to construct a weighted input matrix, which will be used as the basis for deep model training in step S50.

[0045] Step S50: Construct a weighted input matrix X * according to the optimized particle feature subset S * and the optimized weight set W final , input the weighted input matrix X final into the preset double-layer electricity price prediction model, and the model outputs the final predicted electricity price.

[0046] It should be noted that in step S50, the structure of the double-layer electricity price prediction model includes an input layer, a first long short-term memory network layer, a second long short-term memory network layer, a fully connected layer, and an output layer. The first long short-term memory network layer is used to extract short-term dynamic dependencies from the weighted feature matrix; the second long short-term memory network layer is used to learn long-term trend information in the electricity price time series; wherein the second long short-term memory network layer and the single-layer electricity price prediction model structure trained in step S30 are consistent, and the model parameters trained in step S30 are used as the initial parameters of the second long short-term memory network, to realize the migration of the feature subset in the lightweight training stage and the deep modeling stage.

[0047] It can be understood that the double-layer electricity price prediction model adopts a stacked LSTM structure, which divides the time series modeling task into two levels in design: short-term capture and long-term trend extraction. The first layer LSTM processes short-period changes in the original input (such as daily load fluctuations), and the second layer LSTM processes longer time sequence correlations across days and weeks, and the overall structure improves the model's ability to perceive multi-scale price dynamics.

[0048] It should be understood that since the second layer LSTM parameter structure is consistent with the single-layer prediction model trained in step S30, its trained parameters can be directly migrated to the second layer as the initial state, which not only retains the basic time series features learned in the feature selection process, but also avoids the parameter instability problem caused by training from scratch.

[0049] Embodiment two: In addition, the present application provides a high-precision electricity price prediction method system based on a multi-source data fusion algorithm, which adopts the high-precision electricity price prediction method based on a multi-source data fusion algorithm in the above embodiment, and can solve the technical problem of a high-precision electricity price prediction method based on a multi-source data fusion algorithm. Compared with the prior art, the beneficial effects of the high-precision electricity price prediction method system based on a multi-source data fusion algorithm provided by the present application are the same as those of the high-precision electricity price prediction method based on a multi-source data fusion algorithm provided by the above embodiment, and other technical features of the high-precision electricity price prediction method system based on a multi-source data fusion algorithm are the same as those disclosed in the above embodiment method. The features are not repeated here.

[0050] Embodiment three: The present application provides a high-precision electricity price prediction method device based on a multi-source data fusion algorithm, please refer to Figure 2A high-precision electricity price prediction method device based on a multi-source data fusion algorithm includes at least one processor, and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the high-precision electricity price prediction method based on the multi-source data fusion algorithm in the above-mentioned embodiment one. The high-precision electricity price prediction method device based on the multi-source data fusion algorithm in the embodiment of the present application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. The high-precision electricity price prediction method device based on the multi-source data fusion algorithm is only an example, and should not bring any limitation to the function and use range of the embodiment of the present application. The high-precision electricity price prediction method device based on the multi-source data fusion algorithm can include a processing device 1001 (such as a central processor, a graphics processor, or the like), which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 1002 or loaded from a storage device 1003 to a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for the operation of the high-precision electricity price prediction method device based on the multi-source data fusion algorithm are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, and the like; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, and the like; a storage device 1003 including, for example, a magnetic tape, a hard disk, and the like; and a communication device 1009. The communication device 1009 can allow the high-precision electricity price prediction method device based on the multi-source data fusion algorithm to communicate with other devices wirelessly or by wire to exchange data. Although the high-precision electricity price prediction method device based on the multi-source data fusion algorithm with various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or fewer systems can be alternatively implemented or possessed.

[0051] Embodiment four: the application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of a high-precision electricity price prediction method based on a multi-source data fusion algorithm as described above. The computer program product provided by the application can solve the technical problem of a high-precision electricity price prediction method based on a multi-source data fusion algorithm. Compared with the prior art, the beneficial effects of the computer program product provided by the application are the same as those of the high-precision electricity price prediction method based on a multi-source data fusion algorithm provided by the above-mentioned embodiments, and are not described here.

[0052] In particular, according to the embodiments disclosed by the application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed by the application include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by a processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed by the application are executed.

[0053] It should be understood that various parts of the application disclosed can be realized with hardware, software, firmware or a combination thereof. In the description of the above-mentioned embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0054] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalent technologies, the application also intends to include these modifications and variations.

Claims

1. A high-precision electricity price prediction method based on a multi-source data fusion algorithm, characterized in that: Methods include: Step S10: Collect and pre-process the multi-source time series data associated with the power market and the energy storage system, and divide the multi-source time series data into three types of feature cabin data according to their physical sources. The three types of feature cabins include supply feature cabin data G, demand feature cabin data D, and energy storage feature cabin data S; normalize the three types of feature cabin data to the interval [0, 1] and construct the supply feature matrix X G , demand feature matrix X D and energy storage characteristic matrix X S ; Step S20: Introduce binary and real-valued hybrid particle swarm update mechanism, based on the supply feature matrix X G , demand feature matrix X D and energy storage characteristic matrix X S Each particle in the particle swarm is encoded into a two-level structure, and the feature subset of particle i is set according to the constructed two-level structure. And output the two-level structure particle encoding results; Step S30: Based on feature subset Extract the corresponding feature matrix And input it into the preset single-layer electricity price prediction model for training, and calculate the mean absolute percentage error Constructing a fitness function Step S40: Based on the fitness function Iteratively update the two-level structure encoding of each particle and output the final optimized particle feature subset S * and optimize the weight set W * ; Step S50: Optimize the particle feature subset S * and optimize the weight set W * Construct the weighted input matrix X final , the weighted input matrix X final Input into the preset two-layer electricity price prediction model, and the model outputs the final predicted electricity price.

2. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S10, the supply characteristic cabin data G includes wind power output data, photovoltaic output data and thermal power planned output data; the demand characteristic cabin data D includes total load data, temperature data and working day identification data; the energy storage characteristic cabin data S includes energy storage charge state data and charging and discharging power data.

3. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S20, based on the supply feature matrix X G , demand feature matrix X D and energy storage characteristic matrix X S Each particle in the particle swarm is coded as a two-level structure, and the two-level structure coding of particle i is Among them, x i is the two-level structure code of particle i, An indicator value indicating whether particle i uses the preset supply cabin, demand cabin, or energy storage cabin; is the selection position of particle i for the features in each compartment; W i is the weight of the continuous features of particle i in each cabin; || represents the vector splicing operation; where, {0,1} 3 Represents a vector of length 3, each bit has a value of 0 or 1, b G To supply the outer layer switch, b D is the outer layer switch of the demand cabin, b S For the energy storage cabin outer switch, only when a cabin outer switch b (·) =1, the cabin feature bit will take effect.

4. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S40, based on the fitness function Iteratively update the two-level structure encoding of each particle and output the final optimal particle feature subset S * and weight set W * The steps specifically include: updating the outer cabin selection position and the inner feature selection position in the two-level structure using Sigmoid combined with random sampling method; updating the continuous feature weights in the two-level structure using speed and position update formulas; recalculating the prediction error after each round of update and updating the individual optimal and global optimal particle positions, iterating until the fitness function converges or reaches the preset maximum number of iterations, and outputting the final optimal particle feature subset S * and weight set W * .

5. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S30, a preset single-layer electricity price prediction model is constructed based on the long short-term memory network method.

6. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S50, the two-layer electricity price prediction model adopts the comprehensive loss function during the training process. for: Where T is the total length of the training sample; τ is the prediction time step; P(t+τ) is the electricity price at the corresponding time in the preset verification set at time t+τ; is the electricity price at time t+τ predicted by the model; β is the variance regularization penalty coefficient, which is used to suppress prediction fluctuations; Var is the variance of the predicted electricity price, which is used to control the prediction stability.

7. The high-precision electricity price prediction method based on a multi-source data fusion algorithm according to claim 1, characterized in that: In step S50, the structure of the two-layer electricity price prediction model includes an input layer, a first long short-term memory network layer, a second long short-term memory network layer, a fully connected layer, and an output layer. The first long short-term memory network layer is used to extract short-term dynamic dependencies from the weighted feature matrix; the second long short-term memory network layer is used to learn long-term trend information in the electricity price time series. Among them, the second-layer long short-term memory network layer is consistent with the single-layer electricity price prediction model structure trained in step S30, and the model parameters trained in step S30 are used as the initial parameters of the second-layer long short-term memory network to realize the migration of feature subsets in the lightweight training stage and the deep modeling stage.

8. A high-precision electricity price prediction method system based on a multi-source data fusion algorithm, applied to a high-precision electricity price prediction method based on a multi-source data fusion algorithm according to any one of claims 1 to 7, characterized in that: The high-precision electricity price prediction method system based on the multi-source data fusion algorithm includes: The feature preprocessing module is used to collect and preprocess the multi-source time series data associated with the power market and the energy storage system. The multi-source time series data is divided into three types of feature cabin data according to their physical sources. The three types of feature cabins include supply feature cabin data G, demand feature cabin data D, and energy storage feature cabin data S. The three types of feature cabin data are normalized to the interval [0, 1] and the supply feature matrix X is constructed. G , demand feature matrix X D and energy storage characteristic matrix X S ; The two-level structure particle encoding module is used to introduce a binary and real-valued hybrid particle swarm update mechanism based on the supply feature matrix X G , demand feature matrix X D and energy storage characteristic matrix X S Each particle in the particle swarm is encoded into a two-level structure, and the feature subset of particle i is set according to the constructed two-level structure. And output the two-level structure particle encoding results; Single-layer prediction fitness evaluation module is used to evaluate the performance of a feature subset Extract the corresponding feature matrix And input it into the preset single-layer electricity price prediction model for training, and calculate the mean absolute percentage error Constructing a fitness function Particle swarm iterative optimization module is used to optimize the Iteratively update the two-level structure encoding of each particle and output the final optimized particle feature subset S * and optimize the weight set W * ; The two-layer prediction deployment module is used to optimize the particle feature subset S * and optimize the weight set W * Construct the weighted input matrix X final , the weighted input matrix X final Input into the preset two-layer electricity price prediction model, and the model outputs the final predicted electricity price.

9. A high-precision electricity price prediction method and device based on a multi-source data fusion algorithm, characterized in that: The high-precision electricity price prediction method device based on a multi-source data fusion algorithm includes: a memory, a processor, and a high-precision electricity price prediction method program based on a multi-source data fusion algorithm stored in the memory and executable on the processor. When the high-precision electricity price prediction method program based on a multi-source data fusion algorithm is executed by the processor, a high-precision electricity price prediction method based on a multi-source data fusion algorithm according to any one of claims 1 to 7 is implemented.

10. A computer program product, characterized in that The computer program product includes a high-precision electricity price prediction method program based on a multi-source data fusion algorithm. When the high-precision electricity price prediction method program based on a multi-source data fusion algorithm is executed by a processor, a high-precision electricity price prediction method based on a multi-source data fusion algorithm according to any one of claims 1 to 7 is implemented.

Images