Transformer dynamic load capacity prediction method considering flexibility of salt lake chemical load

Abstract

The transformer dynamic load capacity prediction method considering the flexibility of salt lake chemical load comprises: extracting the key influence factors of the adjustable space of the salt lake chemical potassium-lithium-magnesium isomeric load by means of the synergistic continuous neural dynamics and the swarm intelligence algorithm; predicting the time sequence result of the adjustable space of the salt lake chemical potassium-lithium-magnesium isomeric load in the future by means of adaptively fusing the dynamic and static combined space-time characteristics of the salt lake chemical through a gating mechanism; and constructing and training a long short-term memory network for predicting the load capacity change amount of the transformer at a future time according to the transformer operation data before and after the historical salt lake chemical load response, and combining the time sequence result of the adjustable space at a future time obtained by prediction with the long short-term memory network after training to predict the load capacity change amount of the transformer at the future time. The method can effectively improve the accuracy of quantifying the relationship between the adjustable space of the salt lake chemical potassium-lithium-magnesium isomeric load and the dynamic load capacity improvement of the transformer in the park.

Description

Technical Field

[0001] This invention relates to the field of power distribution system control technology, and in particular to a method for predicting the dynamic load capacity of transformers that takes into account the load flexibility of salt lake chemical plants. Background Technology

[0002] Equipment such as slurry preparation tanks, cold crystallizers, and lithium precipitation reactors in salt lake chemical plants serve as buffers for material storage and brine storage due to reaction residence time. Small fluctuations in process conditions, such as supersaturation, are permissible, allowing for flexible adjustment. However, the complex topologies of potassium, lithium, and magnesium extraction processes, involving high-dimensional variables and strong coupling between them, along with inherent material and equipment properties like brine specific heat capacity and electrolytic cell thermal inertia, result in significant time-delay transmission characteristics between operating parameters and process variables, making it difficult to fully exploit the load flexibility of salt lake chemical plants. Against this backdrop, effectively tapping into the flexible adjustment potential of high-energy-consuming loads in salt lake chemical plants and quantifying the relationship between this flexible adjustment potential and the improvement of transformer dynamic load capacity are critical issues that urgently need to be addressed.

[0003] Current load flexibility assessments primarily rely on establishing mathematical models of flexibility such as virtual energy storage, flexibility assessment optimization models, and historical data-driven models. However, due to the complexity of salt lake chemical processes, the high dimensionality of variables, and the strong nonlinear coupling time delay characteristics, the computational difficulty of flexibility mathematical modeling and assessment optimization models is extremely high. Single data-driven methods lack interpretability of results and struggle to accurately capture potential coupling relationships between variables and process time delay characteristics. Furthermore, there is a technological gap in utilizing the adjustable potential of load flexibility to suppress transformer winding hotspot temperature surges, and research on the mapping relationship between load flexibility and transformer load-carrying capacity improvement is lacking. Therefore, there is an urgent need to research a dynamic load capacity prediction method for transformers that considers the flexibility of high-energy-consuming load clusters in salt lake chemical plants. This would enable accurate quantification of the transformer's potential load-carrying capacity improvement, providing a basis for transformer overload risk assessment and equipment expansion planning in industrial parks. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the dynamic load capacity of transformers that takes into account the load flexibility of salt lake chemical industry, so as to improve the accuracy of quantifying the relationship between the flexible and adjustable space of potassium, lithium and magnesium heterogeneous loads in salt lake chemical industry and the improvement of the dynamic load capacity of transformers in the park.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical method: a method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry, comprising: Step S1: Establish a function relating the correlation and redundancy between the "selectable influencing factors of adjustable space" and the "target adjustable space" of potassium, lithium and magnesium heterogeneous load in salt lake chemical industry. Solve this function using continuous neural dynamics and swarm intelligence algorithms to obtain the key influencing factors of adjustable space of potassium, lithium and magnesium heterogeneous load in salt lake chemical industry. Step S2: Based on the key influencing factors, construct a prior knowledge graph of salt lake chemical industry that characterizes the whole process of potassium, lithium and magnesium extraction. Use the historical data of key influencing factors and adjustable space and the prior knowledge graph as input to construct a rolling time-varying graph of salt lake chemical industry. Through a gating mechanism, adaptively fuse the dynamic and static spatiotemporal characteristics of salt lake chemical industry to predict the future time series results of the adjustable space of potassium, lithium and magnesium heterogeneous load of salt lake chemical industry. Step S3: Based on the historical operating data of the transformer before and after the load response of the salt lake chemical plant, construct and train a long short-term memory network to predict the change in transformer load capacity at future moments. Combine the time series results of the adjustable space at a certain future moment obtained in step S2 with the trained long short-term memory network to obtain the prediction result of the change in transformer load capacity at that future moment.

[0006] Preferably, further detailed descriptions of steps S1-S3 in this invention can be found in the following detailed embodiments.

[0007] In another aspect, the present invention also provides a transformer dynamic load capacity prediction system that takes into account the load flexibility of salt lake chemical industry. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned transformer dynamic load capacity prediction method that takes into account the load flexibility of salt lake chemical industry.

[0008] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting the dynamic load capacity of a transformer that takes into account the load flexibility of a salt lake chemical industry.

[0009] In another aspect, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for predicting the dynamic load capacity of transformers that takes into account the load flexibility of salt lake chemical plants.

[0010] The beneficial effects of this invention are: (1) This invention constructs a key influencing factor extraction model for the adjustable space of potassium, lithium and magnesium heterogeneous loads in salt lake chemical industry driven by collaborative continuous neural dynamics and swarm intelligence algorithm. It obtains feature sets such as power of industrial park transformer, output quality, salt dust concentration and supersaturation. On this basis, it constructs a prior knowledge graph of salt lake chemical industry to characterize the mechanism of potassium, lithium and magnesium extraction process. Using the historical data of key influencing factors and adjustable space and the prior knowledge graph as input, it constructs a rolling time-varying graph of salt lake chemical industry to characterize the process time delay of characteristic variables formed by brine cascade utilization, material storage pool buffer and potassium, lithium and magnesium co-production. Through the adaptive fusion of dynamic and static spatiotemporal features of salt lake chemical industry by gating mechanism, it proposes a future time series prediction method for the adjustable space of potassium, lithium and magnesium heterogeneous loads in salt lake chemical industry based on dynamic and static graph fusion transformer, and realizes accurate prediction of the adjustable space of potassium, lithium and magnesium heterogeneous load cluster.

[0011] (2) This invention constructs a historical load adjustment sample consisting of load change, temperature rise suppression, and load capacity improvement by subtracting the transformer measurement data before and after the load response of the salt lake chemical industry. Then, using the transformer's initial operating state, load adjustable space, response duration, etc. as constraints, it guides the ordinary differential equation diffusion model to quickly generate candidate samples of transformer load adjustment that characterize the coupling relationship between load flexibility call-transformer temperature rise suppression-load capacity recovery. These samples are then input into a long short-term memory network, and effective samples are selected by integrating the sample's own characteristics and trial prediction errors. Finally, the long short-term memory network is retrained using the expanded load adjustment sample data, and the load flexibility space is called to suppress the temperature rise in the potential overload range of the transformer, thereby realizing the accurate prediction of the increase in the transformer's dynamic load capacity from load flexibility. Attached Figure Description

[0012] Figure 1 This is a flowchart of the transformer dynamic load capacity prediction method that takes into account the load flexibility of salt lake chemical industry involved in the present invention. Figure 2 This is a comparison chart of the predicted load capacity of transformers in the salt lake chemical industrial park obtained under three different schemes in this embodiment of the invention. Detailed Implementation

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0014] like Figure 1 As shown, the method for predicting the dynamic load capacity of transformers that takes into account the load flexibility of salt lake chemical industry provided by the present invention mainly includes the following steps.

[0015] Step S1: Construct a model for extracting key influencing factors of adjustable spatial load of potassium, lithium and magnesium in salt lake chemical industry driven by collaborative continuous neural dynamics and swarm intelligence algorithms.

[0016] Step S101 involves normalizing environmental data such as salt dust concentration, temperature, humidity, and wind speed collected by the meteorological station, SCADA, and production MES system of the salt lake chemical industrial park; environmental protection data such as pollutant emissions and carbon emissions; production time-series data such as equipment power-process status-output quality; and the spatial time-series data of adjustable load in the salt lake chemical industry obtained through statistical methods. This process yields a set of candidate influencing factors for the adjustable spatial load of potassium, lithium, and magnesium heterogeneous components in the salt lake chemical industry. and load target adjustable spatial time series dataset ,in, For the first One candidate influencing factor The normalized value at time 10:00 N The number of potential influencing factors. , and They represent salt lake chemicals. The baseline load, upward adjustment range, and downward adjustment range at any given time. For the salt lake chemical industry load cluster Individual adjustable spatial time series data in The normalized value at time 10:00 M The number of adjustable spaces is the target. M =3, including the baseline load, upward adjustment space, and downward adjustment space. L The time series length of the time series data.

[0017] Step S102, in order to fully capture the causal relationship between the candidate influencing factors of the adjustable space of potassium, lithium and magnesium loads in salt lake chemical industry and the target adjustable space of salt lake chemical industry load, adopts... The correlation and redundancy of the two factors are calculated using the nearest neighbor Copula transfer entropy, and based on this, a biconvex optimization function for selecting key influencing factors in salt lake chemical industry is established. As shown in the following formula: (1) (2) (3) In equation (1), and These are the characteristic correlation vectors and redundancy matrices composed of the candidate influencing factors and the adjustable space of the target load for salt lake chemical industry. for and Transfer entropy between them For the adjustable load space of salt lake chemical industry One candidate influencing factor With the One candidate influencing factor Transfer entropy between them; Represent real numbers; and These are the key influencing factors selection vectors and auxiliary vectors for the adjustable space of the salt lake chemical load target, respectively. ; , , These are the weights, a sufficiently large penalty parameter, and the number of key influencing factors extracted, respectively. Represents a vector of all 1s; superscript T Indicates transpose; This represents the sum of squares of the components of the Hadamard product of the key influencing factors selection vector and auxiliary vector for the load adjustable space. In equation (2), Indicates from +1 to + of At that moment; For the influencing factors to be selected Time before First lag value, used to capture candidate influencing factors Adjustable space for load target The lagged effects; For load adjustable space in Time before First hysteresis value, used to capture the adjustable space of the load target. Autocorrelation; In equation (3), It is the digamma function; K The nearest neighbor number; S The number of samples for spatially predictable adjustable loads in salt lake chemical industry; To fall into The number of nearest neighbors in the dimensional feature space. They represent falling into , , The number of nearest neighbors in the dimensional feature space; To consider the number of candidate influencing factors and target adjustable space variables for the adjustable space of salt lake chemical load with lag order, and to satisfy... ; Representation Definition This is a concatenation vector of candidate influencing factors and target adjustable space for load adjustable space, taking into account the lag order.

[0018] Step S103: To screen key influencing factors for load-adjustable space, a 3-layer projection neural network is used to construct the following formula. , and A multi-gradient driven continuous neural dynamics model is used to solve the biconvex optimization function. Furthermore, by collaborating with swarm intelligence algorithms to update the convergent optimal solution at each layer in a timely manner to achieve the global optimum, the key influencing factors of the adjustable space of potassium, lithium and magnesium heterogeneous loads in salt lake chemical industry were obtained, as shown in Table 1.

[0019] (4) In the formula, and It is a positive time constant; , , They are respectively , , The derivative with respect to time; Select the inner product of vectors related to the spatial characteristics of adjustable load in salt lake chemical industry and key influencing factors; , and These are a matrix of all ones and an N-order identity matrix, respectively. ; This is the projection operator.

[0020] Table 1 Key Influencing Factors on the Adjustable Loading Space of Potassium, Lithium, and Magnesium Heterogeneous Loads in Salt Lake Chemical Industry ; Step S2: Construct a spatial future time series prediction method for potassium, lithium and magnesium heterogeneous loads in salt lake chemical industry based on dynamic and static graph fusion transformer.

[0021] Step S201: Construct a prior knowledge graph for salt lake chemical industry and extract the static spatiotemporal characteristics of key influencing factors of adjustable load space.

[0022] 1) The 29 key influencing factors of the adjustable spatial load of potassium, lithium and magnesium heterogeneous processes in salt lake chemical industry were categorized according to meteorology, environmental protection, brine supply, potassium extraction section, lithium extraction section, magnesium extraction section and production plan. The time series data were then normalized to obtain the set of key influencing factors of the adjustable spatial load of potassium, lithium and magnesium heterogeneous processes in salt lake chemical industry. ,in For meteorological variables, respectively represent Temperature, humidity, radiance, evaporation, and salt dust concentration at any given time; For environmental variables, respectively represent The green electricity consumption rate, carbon emission intensity, and pollutant emission margin at any given time; The variables for brine supply are respectively represented as follows: The power of the potassium-rich brine conveying pump, the power of the magnesium-rich brine conveying pump, and the power of the main brine conveying pump at all times; For the potassium extraction section, the variables are respectively represented as At any given time, the slurry density, foam height, supersaturation, and moisture content of the wet potassium salt concentrate; For the lithium extraction section, the variables are respectively represented as At any given time, the Li+ outlet concentration, membrane flux, reaction pH, and filter cake moisture content were measured. For the magnesium extraction section, the variables are respectively represented as The impurity concentration, dehydration temperature, bath temperature, and liquid magnesium casting temperature at any given time; For production planning variables, respectively represent The power and output of each factory at any given time.

[0023] 2) Using the key influencing factors of the adjustable space of potassium, lithium and magnesium heterogeneous loading in salt lake chemical industry as graph nodes, construct a prior knowledge graph of salt lake chemical industry based on whether there are process causal relationships between nodes. ,in This is a set of key influencing factors for load-adjustable space. It is a set of directed edges determined by the salt lake chemical process mechanism, material conduction, and energy consumption mapping. The static adjacency matrix is ​​as follows: (8) (9) Equation (8) is a block-level structure representation of the static adjacency matrix of the prior knowledge graph of salt lake chemical industry. The row and column order is based on the above 7 classification order, which illustrates the influence relationship between meteorology, environmental protection, brine supply, potassium extraction section, lithium extraction section, magnesium extraction section and production plan. Equation (9) uses the extraction section and brine supply as examples to reflect the process mechanism of potassium, lithium, and magnesium extraction, in which... , , These are the internal matrices for the potassium extraction, lithium extraction, and magnesium extraction sections, respectively. , , , These are the internal matrix of brine supply and the matrix showing the effect of brine supply on the potassium, lithium, and magnesium extraction processes, respectively.

[0024] 3) Integrate key influencing factors of salt lake chemical industry As input variables and the static adjacency matrix of the prior knowledge graph They are mapped together to the multi-head hidden layer subspace for self-attention aggregation, and a temporal embedding vector is introduced. The static spatiotemporal characteristics of key influencing factors of adjustable load space are extracted through nonlinear mapping.

[0025] (10) (11) (12) In the formula, , , These are the correlation characteristics, static spatial characteristics, and static spatiotemporal characteristics of key influencing factors for the adjustable spatial load of salt lake chemical industry output from the prior knowledge graph. For concatenation functions, The associated features are the output of the first attention head. The number of heads in the attention layer of the static graph encoder for salt lake chemical industry; and These are the failure function and activation function, respectively, representing the correlation characteristics of key influencing factors in the load adjustable space. This is a normalization function for attention-related scores; This represents a feedforward neural network; and This is the static spatial feature weight matrix; This represents the linear projection matrix of the first attention head in the attention layer of the static graph encoder's associated feature extraction layer; The linear projection matrix of the network layer for spatiotemporal feature extraction in a static graph encoder; To output the projection matrix; , This is the bias parameter.

[0026] Step S202: Construct a rolling time-varying diagram of salt lake chemical industry and extract the dynamic spatiotemporal characteristics of key influencing factors of adjustable load space.

[0027] 1) In order to capture the strong coupling and time-delay transmission characteristics among variables in salt lake chemical industry, a query vector is obtained by linearly projecting the key influencing factors of the adjustable load space. Key vector Value vector and query vector Scroll along the time axis after transposition At each time step, a self-attention mechanism is used to extract the cross-time step dynamic evolution correlations between heterogeneous variables, generating a dynamic graph containing dynamic time delay causality. .

[0028] (13) (14) (15) (16) In the formula, Key influencing factors for load adjustable space (rolling number) The transpose query vector at each time step; , , They represent , , The first moment Key influencing factors of adjustable load capacity; They are respectively , The first moment Key influencing factors of adjustable load capacity; They are respectively , The first moment Key influencing factors of adjustable load capacity; For the rolling first Self-attention scores among key influencing factors at each time step; Key influencing factors for load adjustable space (rolling number) +1 time step key vector; For the rolling first The time step The and the first The spatial relationship of key influencing factors of adjustable load space at different times; and These are the node and edge weight sets of the dynamic graph of the salt lake chemical industry, respectively. The weights are determined by the rolling query matrix. The self-attention value is determined at each time step, i.e., the adjacency matrix of the dynamic graph. ; Indicates the scrolling number Dynamic diagram of the time delay of key characteristics of adjustable load space at each time step.

[0029] 2) Graph attention network is used to aggregate spatial dependency information of input variables and temporal embedding vectors are introduced to extract dynamic spatiotemporal features of key influencing factors of adjustable load space.

[0030] (17) In the formula, A dynamic graph is fused to represent key features; For graph attention networks; For the first A mapping matrix for each attention head. To extract the correlation features of key influencing factors of load adjustment space in the dynamic graph multi-head attention output, dynamic spatial features of key influencing factors of load adjustment space in salt lake chemical industry are extracted through temporal embedding and residual feedforward network mapping. With dynamic spatiotemporal characteristics .

[0031] Step S203, adjust the load target adjustable space The data is fed into a transformer to learn its own historical evolution patterns. Combined with a gating mechanism, the dynamic and static spatiotemporal features extracted in parallel by the encoder of the salt lake chemical plant's dynamic and static maps are adaptively fused. The resulting data is then processed through a multi-layer network to output the adjustable spatial future of the potassium, lithium, and magnesium heterogeneous loads in the salt lake chemical plant. The time-series prediction results of the step are as follows: the dynamic and static map encoder extracts the dynamic and static spatiotemporal features of the key influencing factors of the adjustable space of salt lake chemical load in parallel, and uses a gating structure to encode and fuse the dynamic and static spatiotemporal features.

[0032] (18) In the formula, is the spatial dynamic and spatiotemporal characteristic fusion matrix of adjustable load in salt lake chemical industry; g is the dynamic weight allocation of the gated structure. This represents the self-evolutionary hidden layer characteristics of the adjustable historical load space for salt lake chemical industry. For the load-adjustable spatial mask attention-related score normalization function; This is a fusion of the self-evolutionary characteristics of the adjustable space in salt lake chemical industry and the dynamic and static spatiotemporal characteristics of key influencing factors; The linear projection matrix of the load-adjustable spatial mask attention layer; This is the linear projection matrix of the attention layer used to fuse the hidden layer features of the spatially self-evolving load and the dynamic and spatiotemporal features of key influencing factors; These are the weight matrices of a two-layer linear structure of a feedforward neural network (FFN); These are the corresponding bias terms; The mapping matrix for the output layer; For the corresponding bias term; For activation functions; Future adjustable loading space for potassium, lithium, and magnesium heterogeneous components in salt lake chemical industry The time series prediction results of the steps, including future Step reference load Upward adjustment space and downward adjustment space .

[0033] Step S3: Consider the suppression of hotspot temperature surge to predict the dynamic load capacity of transformers in the Salt Lake Chemical Industrial Park.

[0034] Step S301: Construct a non-response baseline trajectory based on the original production plan; perform differential processing and normalization on the historical transformer operating data before and after the load response of the salt lake chemical plant; and construct a curve based on the load adjustment. Transformer hot spot temperature rise Transformer top oil temperature change and load capacity change Historical load adjustment sample , For historical load adjustment sample characteristics in The normalized value at time , include Load adjustment amount at any time Transformer hot spot temperature rise Transformer top oil temperature change and load capacity change In this invention, the temperature rise is the temperature surge of the transformer winding hot spot.

[0035] Historical load regulation samples from salt lake chemical plants are typically small samples, making them prone to overfitting during training. Therefore, this invention selects an ordinary differential equation diffusion model with strong stability and a small sampling step size for data augmentation. Compared to time series generation, this diffusion model is more robust in image reconstruction. Preferably, principal component analysis based on Gram matrix (Gram-PCA) is used to... , , and One-dimensional time-series data is image encoded and mapped to a two-dimensional matrix image.

[0036] (19) (20) (twenty one) (twenty two) In the formula, A and B represent the load regulation, transformer hot spot temperature rise, transformer top oil temperature change, and load capacity change. , Represents a time variable; For the first Characteristics of historical load adjustment samples The normalized value at time; for Time of the first Historical load regulation sample characteristics The corresponding angle value in polar coordinates; Indicates the first Historical load adjustment samples The A characteristic at time and The joint features of the A features at time step A in the angle space, and simultaneously used as a two-dimensional matrix image encoding. In the line, number The pixel values ​​of the column; Indicates the first Historical load adjustment samples The A feature quantity at time t and The joint features of the B features at time step in angle space, and simultaneously used as a two-dimensional matrix image encoding. In the line, number The pixel values ​​of the column; Indicates the first A multi-channel Gram angle field image matrix of historical load regulation samples; For the first A low-dimensional latent space representation of the transformer operating status of a historical load regulation sample; The eigenvector matrix of PCA (PCA is principal component analysis); This is the mean vector of the historical load adjustment sample training set; Represent real numbers; The dimension of the latent space, in this invention, The size is determined by the proportion of variance explained.

[0037] Step S302: Establish and train a long short-term memory network to predict changes in transformer load capacity at future moments.

[0038] 1) Divide the historical load adjustment samples into training set, validation set and test set according to a preset ratio.

[0039] 2) Represent the transformer operating states in the Gram-PCA-processed training set in a low-dimensional latent space. By feeding the ordinary differential equation diffusion model as shown below, a low-dimensional latent space for the transformer's thermal state variables is generated. .

[0040] (6) (7) In the formula, This represents the initial operating state of the transformer in a low-dimensional latent space. The first part is the principal component analysis method based on Gram matrix after processing. A low-dimensional latent space representation of the transformer operating status of a historical load regulation sample; For the first Conditional variables for a historical load adjustment sample; For the diffusion time during the training phase, ; For the first A historical load adjustment sample under condition variables Under the guidance of diffusion time The transformer operating status is represented by a noise-filled latent space. The differential equation of the noise latent space with respect to the diffusion time for the transformer operating state represents the velocity field. ,when Along The sample generation speed is fastest when the direction of the line is straight; The loss function for the diffusion generation model of historical load adjustment samples of transformer load; This represents the total number of historical load adjustment samples. For condition variables Under the guidance of the testing phase, the spread The noise potential space of the transformer operating status at each time step Speed ​​predictor The low-dimensional latent space of the transformer thermal state variables generated by denoising.

[0041] Then Perform inverse principal component transformation and inverse Gram matrix transformation to recover the load regulation candidate samples. .

[0042] 3) Adjust the sample from historical load. Divided into and As input features, The output features are used to train the Long Short-Term Memory network. .

[0043] 4) First, adjust the candidate samples from the load. Divided into and Input into the Long Short-Term Memory network trained by step 3) In the middle, we obtained Predicted changes in transformer load capacity Then With From Divided into The difference is used to obtain the prediction error of the transformer load capacity change. Finally and The sample is sent to the sample selector to filter out effective samples of transformer load regulation that can characterize the coupling relationship between "load flexibility adjustment - transformer temperature rise suppression - load capacity improvement", as shown in the following formula: (twenty three) In the formula, This represents a sample selection probability generator; The prediction error for the transformer load capacity change is the difference between the predicted change in transformer load capacity using a long short-term memory network and the target change in load capacity. The sample retention probability is given by combining the characteristics of the input candidate samples and the subsequent prediction error. ; Indicates the probability of sample retention The 0 / 1 selection mask obtained by Bernoulli sampling, if ,but Identified as a valid sample for transformer load regulation .

[0044] It is worth mentioning that the aforementioned low-dimensional latent space representation of the transformer operating state after Gram-PCA processing... The data is fed into the ordinary differential equation diffusion model to generate a low-dimensional latent space for the transformer's thermal state variables. Then Principal component inverse transformation and Gram matrix inverse transformation are performed to recover the load regulation candidate samples, which are then sent to the sample selector for screening the effective samples of transformer load regulation. As can be seen from formula (6), the ordinary differential equation diffusion model uses the velocity field set by the Rectified Flow. This makes random noise point to The trajectory is close to linear, so the diffusion model of ordinary differential equations can be used to accelerate the generation of effective samples of transformer load regulation within a shorter sampling step.

[0045] To ensure that the effective samples of transformer load regulation generated by the ordinary differential equation diffusion model better reflect the coupling relationship of "load regulation-transformer temperature rise suppression-load capacity improvement" in the salt lake chemical industry, this embodiment introduces the first... Initial transformer load rate within a historical load regulation sample Hotspot temperature Top oil temperature Initial operating status and number of load adjustments Start time End time Duration Equal load regulation characteristics as the first Condition variables for a historical load adjustment sample As shown in the following formula: (twenty four) (25) In the formula, For the first The number of load adjustments within a historical load adjustment sample; , , They represent the first The start time, end time, and duration of load regulation within a historical load regulation sample; , They represent the first The first historical load adjustment sample During the second adjustment time, The change in load at any given time; To set the minimum threshold for load changes and avoid misjudgment of load regulation status due to short-term noise interference and small fluctuations; For the first Conditional variables for a historical load adjustment sample; The first A historical load adjustment sample in The scope for adjusting the load upwards and downwards within a given time period.

[0046] 5) Valid samples of transformer load regulation The load adjustment sample training set is expanded to include the training set. ,use The data was used to retrain the Long Short-Term Memory network. The optimized Long Short-Term Memory network was obtained. .

[0047] Step S303: Based on the predicted potassium, lithium, and magnesium isomer load of the salt lake chemical industry obtained in step S2, the future... Baseline load at time Upward adjustment space and downward adjustment space Determine the potential overload range of the transformer, and within this overload range, aim to achieve the best effect in suppressing the temperature rise of the transformer hot spots and the temperature rise of the top oil layer, and then call... As Load adjustment amount at any time ,Will and Historical load adjustment samples at any given time Sent in together In the middle, the future can be predicted. Change in transformer load capacity at any given time As shown in the following formula: (5).

[0048] On the other hand, based on the same principle as the transformer dynamic load capacity prediction method considering the load flexibility of salt lake chemical industry described in the above embodiments, the present invention also provides a transformer dynamic load capacity prediction system considering the load flexibility of salt lake chemical industry. This system includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the transformer dynamic load capacity prediction method considering the load flexibility of salt lake chemical industry described in the above embodiments. Specifically, the system can be an electronic computer or tablet computer, the processor can be a CPU, GPU, etc., and the memory can be RAM, ROM, EEPROM, CDROM, disk storage medium, or any other medium capable of carrying or storing the computer program and capable of being read by a computer, without limitation herein.

[0049] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned method for predicting the dynamic load capacity of transformers taking into account the load flexibility of salt lake chemical plants. Specifically, the computer-readable storage medium may be RAM, ROM, EEPROM, SSD, CDROM, DVD, USB flash drive, or any other medium capable of carrying or storing the computer program and capable of being read by a computer.

[0050] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for predicting the dynamic load capacity of transformers that takes into account the load flexibility of salt lake chemical plants.

[0051] Following the methods in steps S1-S3 of the aforementioned embodiments, historical operating data from a typical salt lake chemical industrial park in China, including potash, lithium, and magnesium metal manufacturing plants, were selected over the past year and divided into training, validation, and test sets in a 7:2:1 ratio. To verify the effectiveness of the proposed method in predicting the dynamic load capacity of transformers in the industrial park under the condition of considering the load flexibility of salt lake chemical plants, three comparative schemes were set up: Solution 1: The method proposed in this invention; Option 2: Considers the load flexibility of the salt lake chemical industry, but does not consider the suppression of transformer hot spot temperature rise; Option 3: Does not consider the load flexibility of the salt lake chemical industry and the suppression of transformer hot spot temperature rise.

[0052] The predicted load capacity of the transformers in the Salt Lake Chemical Industrial Park under the three schemes are as follows: Figure 1 As shown in Table 2, the minimum load capacity and average load capacity of the transformers in the Salt Lake Chemical Industrial Park were selected as evaluation indicators. The comparison results of the three schemes are shown in Table 2.

[0053] Table 2 Comparison of Prediction Results for Different Schemes ; Depend on Figure 1 It can be seen that the method proposed in this invention maintains a higher and more stable transformer load capacity in all time periods, with a significant increase in peak load. This indicates that considering the load flexibility of the salt lake chemical plant and using it for transformer temperature rise suppression can effectively mitigate the cumulative effect of hot spot temperature rise under heavy overload conditions of the transformer in the main substation of the industrial park, thus shortening the transformer overheating lag recovery time and improving the dynamic load capacity of the equipment. Although Scheme 2 introduces load flexibility and its curve is generally higher than Scheme 3, its load capacity is lower than Scheme 1 due to the cumulative effect of hot spot temperature rise during high load periods. Scheme 3 does not consider load flexibility and temperature rise suppression, resulting in the most severe accumulation of hot spot temperature rise in the transformer, the lowest transformer load capacity trough, and a slower recovery time. As shown in Table 2, Scheme 1 has better minimum and average transformer load capacity than the other two schemes, with a minimum increase of at least 16.73% in minimum load capacity and at least 8.35% in average load capacity. This indicates that the method proposed in this invention can not only improve the upper limit of transformer short-term overload capacity but also improve the average operating capacity throughout the day.

[0054] The above embodiments are preferred implementations of the present invention. In addition, the present invention can be implemented in other ways. Any obvious substitutions without departing from the concept of the present technical solution are within the protection scope of the present invention.

[0055] To facilitate understanding by those skilled in the art of the improvements of this invention over the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this invention.

Claims

1. A method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical plants, characterized in that, include: Step S1: Establish a function relating the correlation and redundancy between the candidate influencing factors and the target adjustable space of potassium-lithium-magnesium heterogeneous loading in salt lake chemical industry. Solve this function using continuous neural dynamics and swarm intelligence algorithms to obtain the key influencing factors of the adjustable space of potassium-lithium-magnesium heterogeneous loading in salt lake chemical industry. Step S2: Based on the key influencing factors, construct a prior knowledge graph of salt lake chemical industry to characterize the mechanism of the entire potassium, lithium and magnesium extraction process, and extract the static spatiotemporal characteristics of key influencing factors with adjustable extraction load. Using historical data and prior knowledge graphs of key influencing factors and adjustable space as input, a rolling time-varying graph of salt lake chemical industry is constructed to extract dynamic spatiotemporal features of key influencing factors of adjustable load space. The adjustable load target space of salt lake chemical industry is fed into transformer to learn its own historical evolution law. Combined with gating mechanism, the dynamic and static spatiotemporal features extracted by the dynamic and static graph encoder of salt lake chemical industry are adaptively fused. The future time series results of the adjustable load space of potassium, lithium and magnesium heterogeneous salt lake chemical industry are predicted through multi-layer network. The dynamic and static graph encoder extracts dynamic and static spatiotemporal features in parallel and uses gating structure to encode and fuse them to obtain dynamic and static spatiotemporal features. Step S3: Based on the historical operating data of the transformer before and after the load response of the salt lake chemical plant, construct and train a long short-term memory network to predict the change in transformer load capacity at future moments. Combine the time series results of the load adjustable space at a certain future moment obtained in step S2 with the trained long short-term memory network to obtain the prediction result of the change in transformer load capacity at that future moment. Step S301: Subtract and normalize the historical transformer operating data before and after the load response of the salt lake chemical plant to construct a data structure. Load adjustment amount at any time Transformer hot spot temperature rise Transformer top oil temperature change and load capacity change Historical load adjustment sample The temperature rise refers to the surge in temperature of the transformer winding hotspots. Then, based on principal component analysis of the Gram matrix, the one-dimensional time-series data in the historical load regulation samples are image-encoded and mapped into a two-dimensional matrix image, yielding a low-dimensional latent space representation of the transformer operating state in the historical load regulation samples. ; Step S302: Construct and train a long short-term memory network to predict future changes in transformer load capacity using historical load adjustment samples. ; Step S303: Based on the future adjustable space of potassium, lithium, and magnesium heterogeneous loading in salt lake chemical industry predicted in step S2... The timing results at any given time determine the potential overload range of the transformer. Within this overload range, the goal is to achieve the best effect in suppressing the temperature rise of transformer hot spots and the temperature rise of top oil. The predicted timing results are then used to determine the optimal time range. The room for adjustment at any time As the load adjustment amount at that moment ,Will and Historical load adjustment samples at any given time Sent in together In the middle, the future can be predicted. Change in transformer load capacity at any given time As shown in the following formula: (5) in, Indicates a time step.

2. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 1, characterized in that: Step S1 includes: Step S101: Collect environmental data, environmental protection data, and production time-series data of the salt lake chemical plant. Calculate the spatial time-series data of the adjustable load of the salt lake chemical plant using statistical methods. Normalize these data to obtain the set of candidate influencing factors for the adjustable spatial load of potassium, lithium, and magnesium heterogeneous components in the salt lake chemical plant. And target adjustable spatial time series dataset ,in, For the first One candidate influencing factor The normalized value at time 10:00 N The number of potential influencing factors. , and They represent salt lake chemicals. The baseline load, upward adjustment range, and downward adjustment range at any given time. For the salt lake chemical industry load cluster Individual adjustable spatial time series data in The normalized value at time 10:00 M The number of adjustable spaces is the target. M =3, including the baseline load, upward adjustment space, and downward adjustment space. L The time series length of the time series data; Step S102, using The nearest neighbor Copula transfer entropy is used to calculate the correlation and redundancy between the candidate influencing factors and the adjustable space of the salt lake chemical industry load target. Based on this, a biconvex optimization function for selecting key influencing factors of the salt lake chemical industry is established. As shown in the following formula: (1) (2) (3) In equation (1), and These are the characteristic correlation vectors and redundancy matrices composed of the candidate influencing factors and the adjustable space of the target load for salt lake chemical industry. for and Transmission entropy between them For the adjustable load space of salt lake chemical industry One candidate influencing factor With the One candidate influencing factor Transfer entropy between them; Represent real numbers; and These are the key influencing factors selection vectors and auxiliary vectors for the adjustable space of the salt lake chemical load target, respectively. ; , , These are the weights, penalty parameters, and the number of key influencing factors extracted, respectively. Represents a vector of all 1s; superscript T Indicates transpose; This represents the sum of squares of the components of the Hadamard product of the key influencing factors selection vector and auxiliary vector for the load adjustable space. In equation (2), Indicates from +1 to + of At that moment; For the influencing factors to be selected Time before First lag value, used to capture candidate influencing factors Adjustable space for load target The lagged effects; For load adjustable space in Time before First hysteresis value, used to capture Autocorrelation; In equation (3), It is the digamma function; K The nearest neighbor number; S The number of samples for spatially predictable adjustable loads in salt lake chemical industry; To fall into The number of nearest neighbors in the dimensional feature space. They represent falling into , , The number of nearest neighbors in the dimensional feature space; To consider the number of candidate influencing factors and target adjustable space variables for the adjustable space of salt lake chemical load with lag order, and to satisfy... ; Representation Definition The concatenation vector of candidate influencing factors and target adjustable space for load adjustable space considering lag order; Step S103, construct the following formula using a 3-layer projection neural network. , and A multi-gradient driven continuous neural dynamics model is used to solve the biconvex optimization function. Furthermore, by coordinating with swarm intelligence algorithms to update the convergent optimal solution at each layer in a timely manner to achieve the global optimum, key influencing factors of the adjustable space of potassium-lithium-magnesium heterogeneous load in salt lake chemical industry were obtained. (4) In the formula, and It is a positive time constant; , , They are respectively , , The derivative with respect to time; Select the inner product of vectors related to the spatial characteristics of adjustable load in salt lake chemical industry and key influencing factors; , and These are a matrix of all ones and an N-order identity matrix, respectively. ; This is the projection operator.

3. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 2, characterized in that: Step S2 includes constructing a prior knowledge graph for salt lake chemical industry and extracting the static spatiotemporal characteristics of key influencing factors for adjustable load space, as follows: 1) The key influencing factors obtained in step S1 are classified according to meteorology, environmental protection, brine supply, potassium extraction section, lithium extraction section, magnesium extraction section and production plan, and their time series data are normalized to obtain the set of key influencing factors for the adjustable spatial load of potassium, lithium and magnesium heterogeneous components in salt lake chemical industry. 2) Using the key influencing factors of the adjustable space of potassium, lithium and magnesium heterogeneous loading in salt lake chemical industry as graph nodes, construct a prior knowledge graph of salt lake chemical industry based on whether there are process causal relationships between nodes. ,in This is a set of key influencing factors for load-adjustable space. It is a set of directed edges determined by the salt lake chemical process mechanism, material conduction, and energy consumption mapping. It is a static adjacency matrix; 3) Use the set of key influencing factors as input variables and the static adjacency matrix of the prior knowledge graph. Self-attention aggregation is performed by mapping to the multi-head hidden layer subspace, and a temporal embedding vector is introduced. Static spatiotemporal features of key influencing factors of load-adjustable space are extracted through nonlinear mapping.

4. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 3, characterized in that: Step S2 further includes constructing a rolling time-varying diagram of salt lake chemical industry and extracting the dynamic spatiotemporal characteristics of key influencing factors of adjustable load space, as follows: 1) Linear projection of key influencing factors of load adjustable space yields query vector, key vector, and value vector, and the query vector is transposed and then scrolled along the time axis. Each time step extracts the cross-time step dynamic evolution correlation between heterogeneous variables through a self-attention mechanism, generating a dynamic graph containing dynamic time delay causality; 2) Graph attention network is used to aggregate spatial dependency information of input variables and temporal embedding vectors are introduced to extract dynamic spatiotemporal features of key influencing factors in adjustable space.

5. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 4, characterized in that: Step S302 includes: 1) Divide the historical load adjustment samples into training set, validation set, and test set according to a preset ratio; 2) Transfer the training set The data is fed into the ordinary differential equation diffusion model to generate a low-dimensional latent space for the transformer's thermal state variables. Then Perform inverse principal component transformation and inverse Gram matrix transformation to recover the load regulation candidate samples. ; 3) Adjust the sample from historical load. Divided into and As input features, The output features are used to train the Long Short-Term Memory network. ; 4) First, adjust the candidate samples from the load. Divided into and Input into the Long Short-Term Memory network trained by step 3) In the middle, we obtained Predicted changes in transformer load capacity Then With From Divided into The difference is used to obtain the prediction error of the transformer load capacity change. Finally and The samples are sent to the sample selector to screen and obtain effective samples of transformer load regulation that characterize the coupling relationship between "load flexibility adjustment - transformer temperature rise suppression - load capacity improvement"; 5) Expand the effective samples of transformer load regulation into the training set, and then retrain using the expanded training set. The optimized Long Short-Term Memory network was obtained. .

6. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 5, characterized in that: The expression for the ordinary differential equation diffusion model is as follows: (6) (7) In the formula, This represents the initial operating state of the transformer in a low-dimensional latent space. The first part is the principal component analysis method based on Gram matrix after processing. A low-dimensional latent space representation of the transformer operating status of a historical load regulation sample; For the first Conditional variables for a historical load adjustment sample; For the diffusion time of the training phase, ; For the first A historical load adjustment sample under condition variables Under the guidance of diffusion time The transformer operating status is represented by a noise-filled latent space. The differential equation of the noise latent space with respect to the diffusion time for the transformer operating state represents the velocity field. ,when Along The sample generation speed is fastest when the direction of the line is straight; The loss function for the diffusion generation model of historical load adjustment samples of transformer load; This represents the total number of historical load adjustment samples. For condition variables Under the guidance of the testing phase, the spread The noise potential space of the transformer operating status at each time step Speed ​​predictor The low-dimensional latent space of the transformer thermal state variables generated by denoising.

7. The method for predicting the dynamic load capacity of transformers considering the load flexibility of salt lake chemical industry according to claim 6, characterized in that: In the expression of the ordinary differential equation diffusion model, the condition variables are the initial operating state of the transformer and the load regulation characteristics. The initial operating state of the transformer includes the initial load rate, hot spot temperature, and top oil temperature. The load regulation characteristics include the number of load regulation cycles, start and end times, and duration.

8. A transformer dynamic load capacity prediction system considering the load flexibility of salt lake chemical industry, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor executes the computer program to implement the transformer dynamic load capacity prediction method considering the load flexibility of salt lake chemical industry as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the method for predicting the dynamic load capacity of a transformer that takes into account the load flexibility of salt lake chemical industry, as described in any one of claims 1-7.

Images