Electric power spot transaction volume dynamic prediction method based on deep learning

Abstract

The invention belongs to the technical field of data analysis and intelligent power transaction management, provides a power spot transaction volume dynamic prediction method based on deep learning, and aims to solve the problem of low accuracy of automatic data analysis of power spot transaction volume prediction in the prior art. According to the method, external dynamic event information influencing the electric power spot transaction is introduced in time to intelligently and automatically fuse more factors influencing the electric power spot transaction, and a first feature vector is extracted based on a large language model under the condition that first type data exists; and determining the second type of data and the second feature vector thereof according to the first feature vector and the second feature vector, predicting the electric power spot transaction volume according to the first feature vector and the second feature vector and based on a pre-trained electric power spot transaction volume probability prediction model, and predicting the electric power spot transaction volume by combining classification prediction, a large language model, a deep learning model, interval prediction and the like. Automatic data analysis based on rich data driving and model driving is realized, and the accuracy of automatic data analysis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis and intelligent power transaction management, and particularly relates to a power spot transaction volume dynamic prediction method based on deep learning. BACKGROUND

[0002] Power spot transaction volume prediction refers to a process of predicting the transaction volume of a power spot market in a specific time period (such as one day, one week or a shorter period) in the future based on historical data, market information, weather conditions, macroeconomics and other factors, and using mathematical models and algorithms. In the field of power market spot transactions, accurate prediction of power transaction volume is the key to ensuring stable operation of the power market and improving resource allocation efficiency.

[0003] In traditional technology, power spot transaction volume prediction is generally performed by collecting multi-dimensional data, such as historical transaction data of the power market (including historical transaction volume, historical electricity price, market participant bidding information, etc.), real-time market data (such as the current market supply and demand status), meteorological data (temperature, humidity, and extreme weather warning information), user electricity demand data (user electricity behavior pattern, demand elasticity), etc., and using corresponding models for prediction.

[0004] However, the inventors have found that traditional power spot transaction volume prediction, which relies on limited multi-dimensional data and uses prediction models to obtain prediction results, cannot fully and accurately capture and adapt to the complex dynamic changes of the power market, resulting in a high possibility of large deviation between the automatic prediction results based on automatic data analysis and the actual transaction situation, loss of accuracy and reliability of automatic data analysis, reduction of the intelligent level of power spot transaction volume prediction, and inability to provide reliable transaction decision basis for market participants, which seriously restricts the intelligent level of scientific management and efficient operation of the power market.

[0005] Therefore, how to improve the accuracy of automatic data analysis of power spot transaction volume prediction has become a problem that needs to be solved in the technical field of data analysis and intelligent power transaction management. SUMMARY

[0006] The technical problem solved by the present application is to improve the accuracy of automatic data analysis of power spot transaction volume prediction in traditional technology.

[0007] To solve the above technical problems, the application provides the following technical scheme: a power spot transaction volume dynamic prediction method based on deep learning, comprising: in response to a selection operation of a preset power spot transaction volume prediction type, determining a target power spot transaction volume prediction type; according to the target power spot transaction volume prediction type and based on a preset data identifier, judging whether there is corresponding first type data, the first type data including at least one of the following latest information related to power: news information, policy and market rule change information, social media public opinion information, analyst report; if the above judgment is yes, obtaining the first type data and extracting a first feature vector corresponding to the first type data based on a feature extraction large language model; determining second type data corresponding to the target power spot transaction volume prediction type and vectorizing the second type data to obtain a second feature vector based on a preset feature vectorization method, the second type data including at least one of the following: resident historical electricity consumption index, industry historical electricity consumption index, power demand index, temperature and humidity index; determining a pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type; according to the first feature vector and the second feature vector and based on the pre-trained power spot transaction volume probability prediction model, predicting the power spot transaction volume to obtain a power spot transaction volume prediction interval corresponding to the target power spot transaction volume prediction type.

[0008] As a preferred scheme of the power spot transaction volume dynamic prediction method based on deep learning, according to the target power spot transaction volume prediction type and based on a preset data identifier, judging whether there is corresponding first type data, comprising: determining a pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type; judging whether the pre-trained power spot transaction volume probability prediction model is set with a preset data identifier; in the case where the preset data identifier is not set, it is determined that there is no corresponding first type data; in the case where the preset data identifier is set, it is judged whether there is target specific content data corresponding to the preset data identifier; in the case where there is no target specific content data, it is determined that there is no corresponding first type data; in the case where there is target specific content data, it is determined that there is corresponding first type data.

[0009] The beneficial effects of this invention are as follows: The method introduces external dynamic event information affecting electricity spot trading in a timely manner, intelligently and automatically integrating more factors affecting electricity spot trading. In the presence of a first type of data, it extracts a first feature vector based on a large language model; determines a second type of data and its second feature vector; and predicts electricity spot trading volume based on the first and second feature vectors and a pre-trained electricity spot trading volume probability prediction model. Thus, for different types of electricity spot trading volume predictions, corresponding pre-trained electricity spot trading volume probability prediction models are set, constructing an adaptive mapping relationship between the electricity spot trading volume prediction type and the pre-trained electricity spot trading volume probability prediction model, which can improve the accuracy of automatic data analysis for electricity spot trading volume prediction. By extracting the first feature vector corresponding to the first type of data based on a feature extraction large language model, it can not only timely introduce external dynamic event information such as news, policy documents, social media sentiment, and analyst reports affecting the electricity market in electricity spot trading volume prediction, but also accurately capture and quantify the key and important aspects of the first type of data by leveraging the powerful language understanding capabilities of the large language model. By combining deep information and a large language model with a pre-trained probabilistic prediction model for electricity spot trading volume based on deep learning, a paradigm shift from "data-driven" to "knowledge and data-driven" dynamic prediction of electricity spot trading volume is achieved. This further improves the accuracy of automatic data analysis in electricity spot trading volume prediction. By constructing a pre-trained probabilistic prediction model for electricity spot trading volume, dynamic interval prediction of electricity spot trading volume is achieved, instead of traditional value prediction (point prediction). Compared to value prediction in traditional techniques, the uncertainty of values ​​corresponding to interval prediction improves the relative certainty of dynamic prediction of electricity spot trading volume. This not only further improves the accuracy of automatic data analysis in electricity spot trading volume prediction but also gives it higher reference value. Therefore, by combining classification prediction, large language models and deep learning models, static and dynamic data, structured and unstructured data, and interval prediction, automatic data analysis based on rich data-driven and model-driven approaches is achieved, improving the accuracy and intelligence level of automatic data analysis in electricity spot trading volume prediction. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the deep learning-based dynamic prediction method for electricity spot trading volume provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the first sub-process of the deep learning-based dynamic prediction method for electricity spot trading volume provided in an embodiment of the present invention. Figure 3A second sub-process schematic diagram of the deep learning-based power spot transaction volume dynamic prediction method provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0011] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments.

[0012] The embodiment of the present application provides a deep learning-based power spot transaction volume dynamic prediction method, which can be applied to devices including but not limited to smart phones, tablet computers, desktop computers, servers, etc., and used when market core participants (such as power generation enterprises, power selling companies, large power users, etc.), market organizers and regulators (such as power transaction centers, power grid dispatching agencies, etc.), market service providers and investors (such as professional energy data analysis and service companies, financial institutions and investors) automatically analyze data to predict power spot transaction volume.

[0013] In the face of the technical problem of low accuracy of automatic data analysis of power spot transaction volume prediction in traditional technology, the inventors propose the deep learning-based power spot transaction volume dynamic prediction method of the embodiment of the present application. The core idea of the embodiment of the present application is: by timely introducing external dynamic event information such as news, policy documents, social media public opinion, analyst reports related to power spot transactions that affect power spot transaction volume, and using the nature of data, the corresponding data is classified (classified as static data and dynamic data, structured data and unstructured data, etc.), to process using the corresponding processing method, at the same time, combined with the following aspects: first, classification, set up the corresponding probability prediction model for different prediction types, realize the focus of the probability prediction model and the self-adaptation of the prediction situation; second, fusion, fuse data of different properties, and fuse large language models and deep learning models, static data and dynamic data, structured data and unstructured data, etc., realize the paradigm shift from "data-driven" to "knowledge and data collaborative driving", for example, based on the fusion of large language models and unstructured data, realize the extraction of deep features that quantitative models are difficult to capture from unstructured text corresponding to external events such as news, policy documents, social media public opinion, analyst reports that affect the power market, market sentiment; third, interval prediction, on the basis of constructing the corresponding probability prediction model, carry out the corresponding interval prediction, no longer carry out value prediction (point prediction), improve the overall prediction certainty through micro-uncertainty, based on this, realize the automatic data analysis based on rich data driving and model driving, improve the accuracy of automatic data analysis, thereby improving the accuracy and intelligence level of automatic data analysis of power spot transaction volume prediction.

[0014] It should be noted that the related terms involved in the embodiments of the present application, although not all include the related words such as "electric power, electric power spot, transaction", as a person skilled in the art can understand, it should be related to the automatic data analysis of the subject of the present application "electric power spot transaction volume dynamic prediction", for example, the "news" involved in the embodiments of the present application refers to the news related to electric power spot transaction, the "policy document" refers to the policy document related to electric power spot transaction, and so on, and the like, which will not be repeated.

[0015] The present application will be described in detail below through specific embodiments.

[0016] Embodiment 1, please refer to Figure 1 , Figure 1 The flowchart of the electric power spot transaction volume dynamic prediction method based on deep learning provided by the embodiments of the present application is shown in the figure. Figure 1 As shown in the figure, in this embodiment, the method includes but is not limited to the following steps S101-S107: S101, in response to the selection operation of the preset electric power spot transaction volume prediction type, determine the target electric power spot transaction volume prediction type.

[0017] Explanatorily, the electric power spot transaction volume prediction type is set in advance, that is, the preset electric power spot transaction volume prediction type, which represents the type of predicting the electric power spot transaction volume, thereby, for different preset electric power spot transaction volume prediction types, the corresponding electric power spot transaction volume prediction data and model prediction method are used to realize the fine mapping and self-adaptation of the preset electric power spot transaction volume prediction type and the prediction method, and to realize the change of the prediction data and the prediction model prediction method used to adapt to the change of the preset electric power spot transaction volume prediction type, for example, the prediction model of different time scales and the data used are very different, such as super short-term prediction more dependent on real-time data, long-term prediction more dependent on macro trend, and so on, thereby, the pertinence and accuracy of electric power spot transaction volume prediction can be improved.

[0018] The preset power spot transaction volume prediction type can be distinguished and set in different dimensions. For example, according to the prediction period (time scale), the prediction can be divided into ultra-short-term prediction (for example, minute level ~ 4 hours), short-term prediction (for example, 1 day ~ 1 week), medium-term prediction (for example, 1 week ~ 1 month), long-term prediction (for example, more than 1 month ~ several years) and the like; according to the spatial granularity (geographical range), the prediction can be divided into node-level prediction (for example, a certain specific power transmission node and the like), regional-level prediction (for example, a pricing region such as a province), and system-level prediction (for example, the entire interconnected power grid and the like); according to the market state (operation characteristics), the prediction can be divided into normal and stable state (for example, supply and demand are relaxed, price fluctuation is small and the like), tense state (for example, supply and demand are balanced, price fluctuation is intensified and the like), extreme state (for example, extreme shortage or excess, price peak or sharp drop and the like), and event-driven state (for example, affected by specific events such as planned events such as unit maintenance and important line commissioning, sudden events such as typhoon and earthquake, and sudden release of policies).

[0019] It can be understood that due to the differences of each of the above-mentioned preset power spot transaction volume prediction types, the prediction data and prediction model and other prediction methods used in the power spot transaction volume prediction can also be different. For example, for the above-mentioned ultra-short-term prediction type, it is difficult to generate external dynamic event information such as news, policy documents, social media public opinion and analyst reports that can affect the power market in a very short time. Therefore, in this case, the corresponding external dynamic event information can not be used, and the like is not repeated here.

[0020] It should be noted that first, the division of the above-mentioned preset power spot transaction volume prediction type is exemplary and is only used to explain and illustrate the different divisions and settings of the power spot transaction volume prediction type, and is not used to make corresponding limitations; second, the division of the preset power spot transaction volume prediction type can be divided in different granularities or units according to different dimensions such as time, space, range, target, object or attribute facing the prediction, as long as each of the preset power spot transaction volume prediction types has certain commonality and can be independently predicted, which is not limited here.

[0021] According to the above idea and setting, when performing power spot transaction volume dynamic prediction, all preset power spot transaction volume prediction types are displayed on the preset prediction type display page, and relevant personnel can determine which type of power spot transaction volume dynamic prediction is performed by selecting the corresponding configuration operation according to the demand. Therefore, in response to the selection operation of the preset power spot transaction volume prediction type, the target power spot transaction volume prediction type is determined, which is the power spot transaction volume prediction type selected by the relevant personnel.

[0022] S102, according to the target power spot transaction volume prediction type, and based on the preset data identifier, it is judged whether there is corresponding first type data, the first type data represents unstructured dynamic data, and the first type data includes at least one of the following latest information related to relevant power: news information, policy and market rule change information, social media public opinion information, analyst report.

[0023] Explanatorily, as described above, for different preset power spot transaction volume prediction types, corresponding prediction data and prediction models are used for prediction methods, thereby, for each preset power spot transaction volume prediction type, the prediction data used is pre-set corresponding data identifier, i.e. preset data identifier, the preset data identifier is used to distinguish different types of data, each preset power spot transaction volume prediction type can use several preset data identifiers corresponding to several types of data, which preset data identifier corresponding data is used by relevant personnel according to the need, which is not limited here. Illustratively, suppose there is an identifier A corresponding to data type A, data type A and identifier A represent A type data, and so on, thereby, there are also identifier B corresponding to data type B, identifier C corresponding to data type C, identifier D corresponding to data type D, based on this, according to the prediction demand, relevant personnel can set that the first preset power spot transaction volume prediction type can use A type data (identifier A represents), B type data (identifier B represents) and C type data (identifier C represents), the second preset power spot transaction volume prediction type can use B type data (identifier B represents), C type data (identifier C represents) and D type data (identifier D represents), and so on, which will not be repeated here.

[0024] Based on the above idea and setting, according to the target power spot transaction volume prediction type, the preset data identifier corresponding to the data type used by the target power spot transaction volume prediction type is determined, and based on the preset data identifier, it is judged whether there is corresponding first type data, two aspects of judgment are needed, first, it is judged whether the target power spot transaction volume prediction type has set to use the first type data corresponding to the data type of "unstructured dynamic data", second, it is judged whether the data of "unstructured dynamic data" has specific, latest information content that needs to be concerned, for example, whether there is relevant news, policy documents, social media public opinion, analyst reports and other external dynamic events that affect the latest specific information content of power spot transaction volume.

[0025] The first type of data represents unstructured dynamic data, the unstructured dynamic data represents information that dynamically changes and is unstructured, and the first type of data is data related to the power spot transaction that affects the power spot transaction. Since the first type of data affects the power spot transaction, there is a correlation and influence relationship between the first type of data and the power spot transaction volume, and specific information is as described above. The "first" involved is only used to distinguish different types of data and does not limit the data. Other similar terms in the embodiments of the present application are used in the same way, and will not be described again. In addition, the unstructured dynamic data represents unstructured, newly generated, and dynamically changing data compared to the original. Generally, it represents information that is constantly updated, highly related to time, and reflects the latest market sentiment and unexpected events. Unstructured dynamic data includes, but is not limited to, external dynamic event information such as news, social media public opinion, analyst reports, and the like that affect the power spot transaction volume and are related to the power spot transaction. For example, unstructured dynamic data includes, but is not limited to, the following data that affects the power spot transaction volume: real-time news about unit tripping, line failure, and natural disasters; temporary power market rule adjustment notifications issued by power exchanges; social media public opinion on real-time discussions and emotional expressions about electricity prices, energy shortages, and electricity on public accounts, micro blogs, and energy professional forums; and analyst comments such as short comments or real-time briefings on the current power market situation issued by financial analysts. Based on this, as a person skilled in the art, based on the research and development capabilities and technical knowledge that they should have, it can be understood that there is a correlation and influence relationship between the first type of data and the power spot transaction volume that conforms to the laws of nature.

[0026] It should be noted and emphasized that the static data and dynamic data, structured data and unstructured data, and other related custom terms involved in the embodiments of the present application are automatically classified and distinguished according to the data properties of the information itself, and are identified and named. The purpose of classification is to subsequently process the corresponding data, and it is not subjective to arbitrarily classify it. The classification does not belong to subjective and arbitrary business rules and intellectual activities. At the same time, although the custom words such as "unstructured dynamic data" belong to intellectual activities in the naming itself, the naming is only used to identify and distinguish the data it corresponds to, and is only used to better understand the data properties and technical solutions. The custom words such as "unstructured dynamic data" can be replaced by other names, and do not affect the corresponding data content and technical solutions of the embodiments of the present application. Just like a person can be called "Zhang San" or "Li Si", but it does not affect the person himself. Other similar terms in the embodiments of the present application are used in the same way, and will not be described again. This should not be the basis for considering the technical solutions of the embodiments of the present application as intellectual activities and business rules.

[0027] S103, if the above judgment is yes, the first type of data is obtained, and the first feature vector corresponding to the first type of data is extracted based on a feature extraction large language model.

[0028] Explanatorily, if the above judgment is yes, i.e. there is corresponding first type of data, the first type of data is obtained, and the first type of data is input into the feature extraction large language model based on the feature extraction large language model, corresponding feature extraction is performed, and the first feature vector corresponding to the first type of data is extracted. For example, relevant text data (such as “notice of requirement to guarantee summer power supply”, “meteorological news predicting continuous high temperature weather next week”, “notice of certain large power plant announcing extension of maintenance” and the like) is input into the feature extraction large language model, and corresponding high-dimensional semantic embedding vector is obtained. Thus, by means of the powerful language understanding ability, reasoning ability and generation ability of the large language model, the first feature vector corresponding to the first type of data is extracted, and in the automatic data analysis of the power spot transaction volume dynamic prediction, the unstructured data corresponding to the external dynamic events such as news, policy documents, social media public opinion and analyst reports that affect the power spot transaction volume are introduced, and by means of the powerful language understanding ability and reasoning ability of the large language model, the key, important and deep information of the first type of data is accurately captured and quantified, thereby helping to improve the accuracy of subsequent automatic data analysis of power spot transaction volume prediction.

[0029] Among them, the feature extraction large language model represents a large language model for semantic information feature extraction. The “feature extraction” involved in the feature extraction large language model is only used to distinguish different large language models from the perspective of use and function, and is not used to limit the large language model. As a person skilled in the art, within the research and development ability and technical cognition that he should have, it can be understood that the large language model (Large Language Model, LLM) represents a huge neural network based on the Transformer architecture and trained on a large amount of text, including but not limited to GPT series, PaLM series, LLaMA series, Ernie, Tongyi Qianwen model. A person skilled in the art can select the corresponding large language model within the optional range according to the needs, and can perform large language model field adaptation training and fine-tuning within the implementable range. The related technical means of the large language model will not be repeated here.

[0030] The first feature vector represents a feature vector based on the semantics of the first type of data. Since the extraction is performed by means of the powerful language processing capabilities of the large language model, such as language understanding and generation capabilities, due to the characteristics of the large language model, the key, important, and deep information of the first type of data can be accurately captured and quantified, thereby obtaining the first feature vector. The first feature vector is in the form of a vector (internal representation / embedding of the large language model), which is the essential form of internal processing and calculation of the large language model. In each stage of processing of the large language model, semantics are represented as vectors (also known as embeddings) in a high-dimensional space. When requesting "embedding" through an API, the output obtained is the vector form. The first feature vector can also be output in the form of information (natural language), which is the final output form of the large language model. The output can be in the corresponding form according to the requirements.

[0031] In S104, the second type of data corresponding to the target power spot transaction volume prediction type is determined, and the second type of data is vectorized based on a preset feature vectorization method to obtain a second feature vector. The second type of data represents structured data, and the second type of data includes at least one of the following: a historical residential electricity consumption index, a historical industrial electricity consumption index, a power demand index, and a temperature and humidity index.

[0032] Explanatorily, the feature vectorization method is pre-set, i.e., a preset feature vectorization method. The preset feature vectorization method represents a method of vectorizing features corresponding to the second type of data. The preset feature vectorization method includes but is not limited to standardization and normalization (applicable to numerical indicators such as temperature and humidity), one-hot encoding and embedding layer (applicable to categorical indicators such as user type, region code, and holiday flag), and periodic encoding (applicable to time indicators such as hour, day of the week, month, and holiday). As a person skilled in the art, based on the technical knowledge and research and development capabilities that they should have, they can select feature vectorization methods including but not limited to the above methods according to requirements, which are not limited herein.

[0033] According to the above concept and setting, the second type of data corresponding to the target power spot transaction volume prediction type is determined, and the second type of data is vectorized based on a preset feature vectorization method to obtain a second feature vector. The second type of data includes at least one of the following: a historical residential electricity consumption index, a historical industrial electricity consumption index, a power demand index, and a temperature and humidity index. As a person skilled in the art, it can be understood that each of the above indicators can be further refined and enriched, which will not be described herein again. Meanwhile, the above indicators affect the power spot transaction, and therefore, there is a correlation and influence relationship between the second type of data and the power spot transaction volume that conforms to the natural law.

[0034] The second type data represents structured data, the structured data refers to the general meaning in the technical field, and will not be repeated here. Similar to the first type data, the second type data is also data related to the power spot transaction, which affects the power spot transaction. Since the second type data also affects the power spot transaction, there is a correlation and influence relationship between the second type data and the power spot transaction volume. Moreover, the second type data includes but is not limited to preset structured dynamic data, preset structured static data, and preset unstructured static data, which will be explained and described below. The second feature vector and the first feature vector have the same data dimension, which facilitates data alignment and fusion.

[0035] Moreover, the preset power spot transaction volume prediction type and the second type data are also corresponding, that is, different preset power spot transaction volume prediction types can use different second type data to realize adaptive and fine mapping between the preset power spot transaction volume prediction type and the second type data, which can further improve the automatic data analysis accuracy and intelligent level of power spot transaction volume prediction.

[0036] S105, determining the pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type.

[0037] Explanatorily, for each type of preset power spot transaction volume prediction type, a corresponding pre-trained power spot transaction volume probability prediction model is set in advance. Different preset power spot transaction volume prediction types generally correspond to different pre-trained power spot transaction volume probability prediction models. Therefore, the pre-trained power spot transaction volume probability prediction model is different with the different preset power spot transaction volume prediction types. The pre-trained power spot transaction volume probability prediction model is adapted to and mapped with the preset power spot transaction volume prediction type, which can improve the automatic data analysis accuracy of the corresponding prediction type power spot transaction volume prediction. Therefore, the pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type is determined.

[0038] The pre-trained electricity spot transaction volume probability prediction model is a pre-trained deep learning model that can output the conditional probability distribution of future electricity spot transaction volume. This model realizes probabilistic prediction, which can quantify the uncertainty of prediction. Based on the distribution, the prediction interval at a specified confidence level (such as 95%) can be generated to represent the range in which the actual future electricity spot transaction volume is most likely to fall. Since general classical deep learning models (such as standard LSTM, CNN, and Transformer) cannot directly output the distribution interval of the prediction value by default, they are usually designed and trained to output a single, deterministic prediction value (point prediction). Therefore, the LSTM, CNN, and Transformer deep learning models need to be improved. Methods such as quantile regression, probability output (direct parameterization of distribution), and ensemble methods can be used to improve the deep learning model, so that the improved pre-trained electricity spot transaction volume probability prediction model can realize probabilistic prediction and output the confidence interval.

[0039] According to the above concept and setting, the pre-trained electricity spot transaction volume probability prediction model corresponding to the target electricity spot transaction volume prediction type is determined, thereby realizing the adaptive mapping and association relationship between the electricity spot transaction volume prediction type and the pre-trained electricity spot transaction volume probability prediction model, focusing on the pre-trained electricity spot transaction volume probability prediction model, and further improving the accuracy of automatic data analysis of electricity spot transaction volume prediction.

[0040] S106, according to the first feature vector and the second feature vector, and based on the pre-trained electricity spot transaction volume probability prediction model, the electricity spot transaction volume is predicted to obtain the electricity spot transaction volume prediction interval corresponding to the target electricity spot transaction volume prediction type.

[0041] Explanatorily, according to the first feature vector and the second feature vector, and based on the pre-trained electricity spot transaction volume probability prediction model, the corresponding electricity spot transaction volume is predicted to obtain the electricity spot transaction volume prediction interval. The electricity spot transaction volume prediction interval represents the range in which the actual future electricity spot transaction volume may fall under a given confidence level (such as 95%).

[0042] S107, in the absence of corresponding first type data, according to the second feature vector, and based on the pre-trained electricity spot transaction volume probability prediction model, the electricity spot transaction volume is predicted to obtain the electricity spot transaction volume prediction interval corresponding to the target electricity spot transaction volume prediction type.

[0043] Illustratively, based on the above description, in the absence of corresponding first type data, it is indicated that there is no external unstructured dynamic event information such as news, policy documents, social media public opinion, analyst reports, etc. affecting the electricity market, and the electricity spot transaction volume is directly predicted according to the second feature vector and based on the pre-trained electricity spot transaction volume probability prediction model, to obtain the electricity spot transaction volume prediction interval.

[0044] The method of the embodiment of the present application can improve the automatic data analysis accuracy of the power spot transaction quantity prediction by timely introducing external dynamic event information affecting the power spot transaction, intelligently and automatically fusing more factors affecting the power spot transaction, and based on a large language model, extracting a first feature vector in the presence of first type data; determining second type data and a second feature vector thereof, and based on the first feature vector and the second feature vector, and based on a pre-trained power spot transaction quantity probability prediction model, predicting the power spot transaction quantity. For different types of power spot transaction quantity prediction, a corresponding pre-trained power spot transaction quantity probability prediction model is set, a mapping adaptive relationship between the power spot transaction quantity prediction type and the pre-trained power spot transaction quantity probability prediction model is constructed, and the automatic data analysis accuracy of the power spot transaction quantity prediction is improved. By extracting the first feature vector corresponding to the first type data based on the feature extraction large language model, not only can external dynamic event information such as news, policy documents, social media public opinion, and analyst reports affecting the power market be introduced in time in the power spot transaction quantity prediction, but also the key, important, and deep information of the first type data can be accurately captured and quantified by means of the powerful language understanding ability of the large language model. The large language model and the pre-trained power spot transaction quantity probability prediction model based on deep learning are combined to realize the paradigm shift from "data-driven" to "knowledge and data collaborative-driven" power spot transaction quantity dynamic prediction, which can further improve the automatic data analysis accuracy of the power spot transaction quantity prediction. By constructing the pre-trained power spot transaction quantity probability prediction model, the interval prediction of the power spot transaction quantity is realized, and the traditional value prediction (point prediction) is no longer needed. Compared with the value prediction in the traditional technology, the relative certainty of the power spot transaction quantity dynamic prediction is improved by the value uncertainty corresponding to the interval prediction, which can further improve the automatic data analysis accuracy of the power spot transaction quantity prediction and give the automatic data analysis of the power spot transaction quantity dynamic prediction higher reference value. Therefore, by combining classification prediction, a large language model, and a deep learning model, static data and dynamic data, structured data and unstructured data, and interval prediction, automatic data analysis based on rich data driving and model driving is realized, the accuracy of the automatic data analysis is improved, the automatic data analysis accuracy and the intelligent level of the power spot transaction quantity prediction are improved, the accuracy of the power spot transaction quantity dynamic prediction is improved, and the corresponding decision-makers of the power spot transaction can be provided with corresponding psychological expectations, risk information, and risk references, thereby improving the automatic level and the intelligent level of the data analysis of the power transaction management.

[0045] In an embodiment, please refer to Figure 2 , Figure 2 The first sub-process diagram of the power spot transaction quantity dynamic prediction method based on deep learning provided by the embodiment of the present application is shown in FIG. 1. Figure 2As shown, in this embodiment, according to the target power spot transaction volume prediction type, and based on the preset data identifier, it is judged whether there is corresponding first type data, including: S201, determining the pre-training power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type; S202, judging whether the pre-training power spot transaction volume probability prediction model is set with a preset data identifier; S203, in the case where the preset data identifier is not set, it is determined that there is no corresponding first type data; S204, in the case where the preset data identifier is set, it is judged whether there is target specific content data corresponding to the preset data identifier; S205, in the case where there is no target specific content data, it is determined that there is no corresponding first type data; S206, in the case where there is target specific content data, it is determined that there is corresponding first type data.

[0046] Explanatorily, as described above, to judge whether there is corresponding first type data, two aspects of judgment are needed, first, to judge whether the target power spot transaction volume prediction type is set with first type data corresponding to “unstructured dynamic data” type data, second, to judge whether “unstructured dynamic data” type data has specific, latest specific information content that needs to be concerned, thus, first, the pre-training power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type is determined; according to the pre-training power spot transaction volume probability prediction model, it is further judged whether the pre-training power spot transaction volume probability prediction model is set with corresponding data, whether the pre-training power spot transaction volume probability prediction model is set with corresponding data is identified through the preset data identifier of the corresponding data; in the case where the preset data identifier is not set, it is indicated that the pre-training power spot transaction volume probability prediction model is not set with corresponding type data, it is determined that there is no corresponding first type data; in the case where the preset data identifier is set, it is indicated that the pre-training power spot transaction volume probability prediction model is set with corresponding type data, it is further judged whether there is target specific content data corresponding to the preset data identifier, the target specific content data represents specific content of corresponding information, for example, news content corresponding to news affecting the power market, file content corresponding to policy files, public opinion content corresponding to social media public opinions, report content corresponding to analyst reports, etc.; in the case where there is no target specific content data, it is determined that there is no corresponding first type data; in the case where there is target specific content data, it is determined that there is corresponding first type data.

[0047] Further, the method further includes: determining whether the pre-trained power spot transaction volume probability prediction model is set with a preset data identifier, including: determining, according to the pre-trained power spot transaction volume probability prediction model, whether a preset data switch corresponding to the preset data identifier is opened; determining that the pre-trained power spot transaction volume probability prediction model is set with the preset data identifier in a case where the preset data switch is opened; and determining that the pre-trained power spot transaction volume probability prediction model is not set with the preset data identifier in a case where the preset data switch is not opened.

[0048] Specifically, a data switch, i.e., a preset data switch, is set in advance, and the preset data switch indicates a switch for whether the pre-trained power spot transaction volume probability prediction model needs to pay attention to, introduce, and use the first type of data for reasoning. In the embodiment of the present application, by default, “the preset data switch is opened, and the pre-trained power spot transaction volume probability prediction model introduces the first type of data; the preset data switch is closed, and the pre-trained power spot transaction volume probability prediction model does not introduce the first type of data”, within a comprehensible range, “the preset data switch is closed, and the pre-trained power spot transaction volume probability prediction model introduces the first type of data; the preset data switch is opened, and the pre-trained power spot transaction volume probability prediction model does not introduce the first type of data” can also be used, and the specific implementation of the preset data switch is not limited in the embodiment of the present application. Thus, the self-learning in the pre-trained power spot transaction volume probability prediction model is combined with the user interaction interface (page), the interaction and intervention of the pre-trained power spot transaction volume probability prediction model are realized, and the selectability (according to the analysis required), flexibility, and accuracy of the automatic data analysis of the power spot transaction volume prediction are improved. An exemplary implementation process includes but is not limited to the following steps: 1) designing a user-friendly interface (such as a webpage or a dashboard) to show the user a preset data identifier corresponding to the first type of data that the pre-trained power spot transaction volume probability prediction model can use, the preset data identifier is associated with a preset data switch, and the preset data switch controls whether the preset data identifier and the first type of data corresponding thereto are introduced or not; 2) responding to the user's configuration operation on the preset data switch, translating the user's abstract operation on the preset data switch in the front end into specific parameters that the pre-trained power spot transaction volume probability prediction model can understand, for example, if the user selects to close the preset data switch, the feature value is directly set to 0 or a null value before the first type of data is sent to the pre-trained power spot transaction volume probability prediction model for reasoning, or the first type of data is not sent to the pre-trained power spot transaction volume probability prediction model for reasoning; and 3) the pre-trained power spot transaction volume probability prediction model performs reasoning according to the setting of step 2).

[0049] According to the above concept and description, when judging whether the pre-training power spot transaction volume probability prediction model is set with the preset data identifier, first, according to the pre-training power spot transaction volume probability prediction model, it is judged whether the preset data switch corresponding to the preset data identifier is opened; in the case that the preset data switch is opened, it is determined that the pre-training power spot transaction volume probability prediction model is set with the preset data identifier; in the case that the preset data switch is not opened, it is determined that the pre-training power spot transaction volume probability prediction model is not set with the preset data identifier.

[0050] Further, judging whether the preset data switch corresponding to the preset data identifier is opened comprises: in response to the start of the power spot transaction volume dynamic prediction corresponding to the pre-training power spot transaction volume probability prediction model, prompting the user whether to open the preset data switch based on a preset prompt page; detecting the configuration operation of the user on the preset data switch; in response to the configuration operation of the user opening the preset data switch, opening the preset data switch; in response to the configuration operation of the user not opening the preset data switch, not opening the preset data switch.

[0051] Specifically, based on the above description of the preset data switch, judging whether the preset data switch corresponding to the preset data identifier is opened comprises: in response to the start of the power spot transaction volume dynamic prediction corresponding to the pre-training power spot transaction volume probability prediction model, prompting the user whether to open the preset data switch based on a preset prompt page; detecting the configuration operation of the user on the preset data switch; in response to the configuration operation of the user opening the preset data switch, opening the preset data switch; in response to the configuration operation of the user not opening the preset data switch, not opening the preset data switch, thereby combining the self-learning inside the pre-training power spot transaction volume probability prediction model with the user interaction interface (page), allowing the user to select whether to open the preset data switch according to the needs, realizing the interaction and intervention of the pre-training power spot transaction volume probability prediction model, and improving the flexibility of automatic data analysis of power spot transaction volume prediction.

[0052] Further, before judging whether there is the target specific content data corresponding to the preset data identifier, it further comprises: based on a preset legal and compliant network information crawling method, monitoring whether there is network information corresponding to the first type of data; in the case that the network information exists, acquiring the network information to obtain the target specific content data.

[0053] Specifically, a legal and compliant network information crawling manner is preset, i.e., a preset legal and compliant network information crawling manner, which represents a legal and compliant network information crawling manner, and legal and compliant represents that relevant data collection meets the requirements of relevant laws and regulations, such as the Personal Information Protection Law of China, the GDPR (General Data Protection Regulation of the European Union), or information security standards of other countries and regions, and the network information crawling manner represents an automatic crawling manner of online information (usually referred to as "web crawler" or "web scraping"), which can refer to existing technical means, and will not be described here.

[0054] According to the above concept and setting, before determining whether the target specific content data corresponding to the preset data identifier exists, it further includes: monitoring whether the network information corresponding to the first type of data exists based on the preset legal and compliant network information crawling manner; and obtaining the network information to obtain the target specific content data in the case that the network information exists, thereby introducing the unstructured data corresponding to the external dynamic events such as news, policy documents, social media public opinion, and analyst reports that affect the power spot trading volume in a timely manner in the dynamic prediction of the power spot trading volume, and improving the accuracy of automatic data analysis of the dynamic prediction of the power spot trading volume.

[0055] By judging whether the pre-trained power spot trading volume probability prediction model needs to fuse the first type of data, the embodiment of the present application can flexibly set the first type of data according to the needs, thereby flexibly determining whether to introduce external dynamic event information such as news, policy documents, social media public opinion, and analyst reports that affect the power market in the prediction of the power spot trading volume, for example, in a shorter prediction time interval of half an hour, in the case that there is no new external dynamic event information, it is not necessary to introduce external dynamic event information, otherwise, external dynamic event information can be selected to be introduced, which can improve the flexibility of automatic data analysis, thereby improving the accuracy of automatic data analysis of the corresponding dynamic prediction of the power spot trading volume, and improving the flexibility and efficiency of automatic data analysis of the corresponding dynamic prediction of the power spot trading volume.

[0056] In an embodiment, determining the second type data corresponding to the target power spot transaction volume prediction type comprises at least one of the following: determining preset structured dynamic data corresponding to the target power spot transaction volume prediction type, wherein the preset structured dynamic data comprises at least one of the following: latest power spot transaction clearing result, latest weather temperature and humidity; determining preset structured static data corresponding to the target power spot transaction volume prediction type, wherein the preset structured static data comprises at least one of the following: historical power spot price, historical power spot transaction volume; determining preset unstructured static data corresponding to the target power spot transaction volume prediction type, wherein the preset unstructured static data comprises historical policy and market rule change information; and accordingly, at least one of the preset structured dynamic data, the preset structured static data and the preset unstructured static data is taken as the second type data.

[0057] Explanatorily, preset structured dynamic data corresponding to each preset power spot transaction volume prediction type is set in advance, which, based on similar understanding of the above unstructured dynamic data, represents structured, newly generated, dynamically changed corresponding data (i.e. data changing with time at high frequency and existing in regular numerical or category form) over time, usually time series data with fixed collection frequency, and the preset structured dynamic data comprises but is not limited to the following data affecting power spot transaction volume: latest power spot transaction clearing result, latest weather temperature and humidity, auxiliary service bidding data and other market data; real-time load (power demand), frequency, regional control error (ACE), real-time and predicted output of wind power / solar power, water and thermal power start-up mode and other system operation data; real-time temperature, humidity, wind speed, sunshine intensity (from weather stations) and other physical data. Based on this, as a person skilled in the art, on the basis of research and development ability and technical cognition that he should have, it can be understood that there is a correlation and influence relationship between the preset structured dynamic data and the power spot transaction volume in accordance with natural laws.

[0058] Similarly, the preset structured static data represents the basic and background structured data that does not change over time or changes at a very low frequency, is usually the attribute of the corresponding physical entity or legal entity, and is the "base" for analyzing the power spot transaction volume prediction problem. The preset structured static data includes but is not limited to the following data that affect the power spot transaction volume: rated capacity of the generator set, fuel type (coal, gas, water), efficiency, minimum technical output; power transmission capacity limit of the power transmission line, impedance; latitude and longitude coordinates of the power plant, substation and load center; part of the long-term signed and fixed annual power purchase agreement (PPA) electricity quantity and price (although the contract itself may expire, it is static within the effective period); division of provinces and cities, division of electricity price zones. Based on this, as a person skilled in the art, on the basis of the research and development capabilities and technical cognition that he or she should have, it can be understood that there is a correlation and influence relationship between the preset structured static data and the power spot transaction volume that conforms to the natural law.

[0059] The preset unstructured static data represents relatively stable and infrequently changing data in an unstructured form, which provides a rule framework and long-term background knowledge for the operation of the power spot trading market. The preset unstructured static data includes but is not limited to the following data that affect the power spot transaction volume: documents such as "Electricity Market Operation Rules" and "Renewable Energy Consumption Guarantee Implementation Scheme"; market design documents such as market manuals and settlement rules long documents published by the power exchange; infrequently updated in-depth technical reports such as long-term development plans of power grids and energy white papers; infrastructure blueprints such as design drawings, technical specifications of power plants and power grids (unchanged unless modified). Based on this, as a person skilled in the art, on the basis of the research and development capabilities and technical cognition that he or she should have, it can be understood that there is a correlation and influence relationship between the preset unstructured static data and the power spot transaction volume that conforms to the natural law.

[0060] It should be noted that the above-mentioned unstructured dynamic data, structured dynamic data, unstructured static data and structured static data are only data classification and categorization from the data storage structure, information form and other data properties of the data itself. It is only used for corresponding processing according to the different storage forms, information forms and other data properties of the data to improve the adaptability and processing quality of data processing and analysis, and does not limit the specific content, does not affect the automatic data analysis of the specific content on the power spot transaction volume prediction, and does not affect the correlation and influence relationship between the specific indicators and the power spot transaction volume prediction result that conforms to the natural law.

[0061] According to the above concept and setting, based on the automatic data analysis required for the power spot transaction volume prediction, the second feature vector corresponding to the target power spot transaction volume prediction type is determined, including at least one of the following: determining the preset structured dynamic data corresponding to the target power spot transaction volume prediction type, the preset structured dynamic data including at least one of the following: the latest power spot transaction clearing result, the latest weather temperature and humidity, the latest power spot transaction clearing result representing the latest "transaction price and transaction list" formed after market competition, and the latest weather temperature and humidity representing the latest weather temperature and humidity; determining the preset structured static data corresponding to the target power spot transaction volume prediction type, the preset structured static data including at least one of the following: historical power spot price, historical power spot transaction volume, representing the price and transaction volume of past power spot transactions; determining the preset unstructured static data corresponding to the target power spot transaction volume prediction type, the preset unstructured static data including historical policy and market rule change information, such as the "Electricity Market Operation Rules" and the "Renewable Energy Consumption Guarantee Implementation Scheme" documents; and at least one of the preset structured dynamic data, the preset structured static data, and the preset unstructured static data is used as the second type data.

[0062] Further, based on the preset feature vectorization method, the second type data is vectorized to obtain a second feature vector, including: according to the preset unstructured static data, and based on a feature extraction large language model, performing corresponding semantic information feature extraction to obtain a corresponding second feature vector.

[0063] Specifically, based on the similar processing of the first feature vector corresponding to the first type data, the preset unstructured static data corresponding to the target power spot transaction volume prediction type is determined; according to the preset unstructured static data, and based on a feature extraction large language model, corresponding semantic information feature extraction is performed to obtain a corresponding second feature vector, thereby, with the powerful language understanding ability and reasoning ability of the large language model, the key, important, and deep information of the preset unstructured static data is accurately captured and quantified, the preset unstructured static data is converted into corresponding structured static data, thereby helping to improve the accuracy of subsequent automatic data analysis of the corresponding power spot transaction volume prediction through data fusion.

[0064] In this embodiment of the invention, by using at least one of preset structured dynamic data, preset structured static data, and preset unstructured static data as the second type of data, data of different dimensions, sources, and types are integrated from the perspective of data properties to perform automatic data analysis for dynamic prediction of electricity spot transactions. By combining multi-dimensional and multi-type information in the automatic data analysis for dynamic prediction of corresponding electricity spot transaction volume, the accuracy of the automatic data analysis for dynamic prediction of corresponding electricity spot transaction volume can be further improved.

[0065] In one embodiment, please refer to Figure 3 , Figure 3 This is a schematic diagram of the second sub-process of the deep learning-based dynamic prediction method for electricity spot trading volume provided in an embodiment of the present invention. Figure 3 As shown, in this embodiment, based on the first feature vector and the second feature vector, and based on the pre-trained electricity spot trading volume probability prediction model, the electricity spot trading volume is predicted to obtain the electricity spot trading volume prediction interval corresponding to the target electricity spot trading volume prediction type. This includes: S301, predicting the electricity spot trading volume based on the first feature vector and the second feature vector, and based on the pre-trained electricity spot trading volume probability prediction model, to obtain an initial electricity spot trading volume prediction interval and its corresponding explanatory feature vector; S302, determining a preset prompt word template corresponding to the pre-trained electricity spot trading volume probability prediction model; S303, embedding the initial electricity spot trading volume prediction interval and the explanatory feature vector into the preset prompt word template to obtain a target prompt word; S304, generating the electricity spot trading volume prediction interval corresponding to the initial electricity spot trading volume prediction interval and its explanatory text based on the target prompt word and a text generation language model.

[0066] Explained, a pre-set prompt word template is used, which is a template based on prompt words. The prompt words (Prompt) represent instructions, questions, or text information passed to the corresponding large language model. The relevant technical means of prompt words can refer to the relevant existing technologies in the field of large language models, which will not be elaborated here.

[0067] According to the above concept and setting, according to the first feature vector and the second feature vector, and based on the pre-trained power spot transaction volume probability prediction model, the power spot transaction volume is predicted to obtain the power spot transaction volume prediction interval corresponding to the target power spot transaction volume prediction type, including: according to the first feature vector and the second feature vector, and based on the pre-trained power spot transaction volume probability prediction model, the corresponding power spot transaction volume is predicted to obtain the initial power spot transaction volume prediction interval based on the deep learning model and the corresponding explanatory feature vector. The initial power spot transaction volume prediction interval represents the corresponding prediction result based on the deep learning model, and the explanatory feature vector represents the vector expression of which features in the input feature sequence that the prediction result is concerned about and based on when the pre-trained power spot transaction volume probability prediction model outputs the prediction result. It should be noted that the pre-trained power spot transaction volume probability prediction model based on the deep learning model, if it is based on a model with an attention mechanism (such as Transformer or TFT model), can directly output its decision basis (such as which input features are concerned about) while making a prediction, that is, directly output the explanatory feature vector. The pre-trained power spot transaction volume probability prediction model based on the deep learning model, if it is based on a deep learning model itself that does not directly provide explanation, such as convolutional neural network (CNN) and multi-layer perceptron (MLP), recurrent neural network (RNN, LSTM, GRU), etc. can use external tools (such as SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations)) to analyze and explain why it makes this decision, that is, with the help of external tools such as SHAP and LIME, the explanatory feature vector is obtained. For example, SHAP measures the importance of each feature value by calculating its contribution to the prediction result, and LIME constructs a simple and interpretable (such as linear) model near the prediction point of the original model, and uses this simple model to approximate the decision logic of the complex model in the local. The use of external tools such as SHAP and LIME can refer to existing technical means, which will not be described here.

[0068] The preset prompt template corresponding to the pre-training power spot transaction volume probability prediction model is determined again; and the initial power spot transaction volume prediction interval and the explanatory feature vector are embedded into the preset prompt template to obtain a target prompt; then, according to the target prompt and based on a text generation large language model, a power spot transaction volume prediction interval corresponding to the initial power spot transaction volume prediction interval and an explanatory text thereof are generated, the power spot transaction volume prediction interval represents the final prediction result, which is consistent with the initial power spot transaction volume prediction interval, both represent the prediction result output by the deep learning model, and different customizations are only used for clear description of the technical solution and convenient understanding of the technical solution, wherein the explanatory text represents a text automatically generated by the text generation large language model to explain and describe the prediction basis corresponding to the power spot transaction volume prediction interval. Therefore, with the help of the large language model, the large language model generally uses structured input (prompt), rich world knowledge (self-learning during the training process of the large language model), and powerful logical reasoning to better explain and describe the prediction result corresponding to the initial power spot transaction volume prediction interval (i.e., the power spot transaction volume prediction interval), understand the reasons behind the prediction result, solve the problem of lack of explainability of the output result of the LSTM, Transformer and other deep learning models due to the “black box” characteristics, increase the automatic data analysis credibility of the power spot transaction volume prediction interval, increase the understandability of the automatic data analysis of the power spot transaction volume prediction interval, break the communication barrier between the AI model and the human expert, establish trust, and enable the prediction result to be quickly converted into a transaction decision, thereby improving the practicality.

[0069] Further, embedding the initial power spot transaction volume prediction interval and the explanatory feature vector into the preset prompt template to obtain a target prompt includes: generating a target query condition based on the initial power spot transaction volume prediction interval and the explanatory feature vector and based on a preset query condition generation manner; performing retrieval enhancement on the target query condition based on a preset retrieval enhancement generation paradigm to obtain target retrieval enhancement data; and embedding the initial power spot transaction volume prediction interval, the explanatory feature vector, and the target retrieval enhancement data into the preset prompt template to obtain the target prompt.

[0070] Specifically, the query condition generation method is preset, that is, the preset query condition generation method, which represents a method of generating a query condition according to an initial power spot transaction volume prediction interval and an explanatory feature vector. The preset query condition generation method includes but is not limited to metadata-based template filling (filling key data in the model output into a predefined query template), and query construction based on feature importance (filtering key features and their values according to feature importance scores (such as SHAP) to construct a query). Among them, the metadata-based template filling means that query templates are designed for different types of explanation requirements in advance, and then the specific values of the model output are filled into these templates; the query construction based on feature importance means directly using SHAP and other feature importance scores to construct a query to ensure that the search content is completely consistent with the model decision (such as including the following steps: sorting by feature importance absolute value; selecting key features with importance higher than a certain threshold; combining the names and current values of the key features into a query).

[0071] Retrieval-Augmented Generation (RAG) is a technique that combines information retrieval (IR) systems with the generation capabilities of large language models (LLMs). As a person skilled in the art, with the research and development capabilities and technical knowledge that they should have, it can be understood that the core idea of RAG is to retrieve relevant information from a pre-set external knowledge base before asking the LLM to answer a question or generate text, and then provide this information as context to the LLM, so that the LLM generates more accurate, relevant, and fact-based answers. The workflow of RAG can be divided into two main stages: retrieval and augmented generation. The relevant content of RAG can be referred to existing corresponding technologies, which will not be described here.

[0072] According to the above idea and setting, the initial power spot transaction volume prediction interval and the explanatory feature vector are embedded into the preset prompt word template to obtain the target prompt word, including: generating a target query condition according to the initial power spot transaction volume prediction interval and the explanatory feature vector, and based on a preset query condition generation mode; according to the target query condition, and based on a preset retrieval enhancement generation paradigm, and based on a pre-set corresponding knowledge base (i.e. the above-mentioned external knowledge base relative to the large language model, the knowledge base content can include: documents containing macro market principles, Basic Principles of Power Supply and Demand Balance, Research Report on Meteorological Conditions on Power Load, Case Analysis of the Impact of Unplanned Unit Shutdown on Market Price, and other knowledge content such as past analyst articles, generally a knowledge base set based on the needs of the enterprise), the target query condition is enhanced to obtain target retrieval enhancement data; the initial power spot transaction volume prediction interval, the explanatory feature vector and the target retrieval enhancement data are embedded into the preset prompt word template to obtain the target prompt word, thereby automatically retrieving relevant information (such as historical similar events, power market rules) from the corresponding knowledge base with the help of RAG, and achieving retrieval enhancement with the help of historical data.

[0073] In the embodiments of the present application, by combining a large language model, the output of the pre-trained power spot transaction volume probability prediction model is automatically explained and described, especially based on retrieval enhancement generation, the corresponding text of explanation and description is automatically generated, not only can the prediction result of the pre-trained power spot transaction volume probability prediction model output have stronger prediction basis and credibility, but also based on retrieval enhancement, corresponding retrieval is performed in the corresponding database, and the corresponding prediction result is verified and supported by historical data, so that the accuracy and credibility of the power spot transaction volume probability prediction automatic data analysis can be further improved through verification.

[0074] It should be noted that the deep learning-based power spot transaction volume dynamic prediction method described in each of the above embodiments can recombine the technical features contained in different embodiments as needed to obtain a combined implementation scheme, but all within the scope of protection required by the present application.

[0075] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (or computer- readable storage media) having computer-usable program code embodied in the medium. The medium can be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-usable storage media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired computer program code in the form of computer-usable program code means which can be accessed by a computer. When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or a combination of hardwired or wireless) to a computer, the computer properly views the connection as a computer-usable storage medium. Thus, any such connection is properly termed a computer-usable storage medium. Combinations of the above should also be included within the scope of the computer-usable storage media. Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or a plurality of blocks.

[0076] The relevant data collection in the embodiments of the present application meets the requirements of relevant laws and regulations, such as the Personal Information Protection Law of China, the GDPR (General Data Protection Regulation of the European Union) or other information security standards in other countries and regions.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalent alternatives without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based dynamic prediction method for power spot transaction volume, characterized in that, The method comprises the following steps: In response to a selection operation of a preset power spot transaction volume prediction type, a target power spot transaction volume prediction type is determined; According to the target power spot transaction volume prediction type and based on a preset data identifier, it is determined whether there is corresponding first type data, the first type data including at least one of the following latest information related to electricity: news information, policy and market rule change information, social media public opinion information, analyst report; if the above determination is yes, the first type data is obtained, and a first feature vector corresponding to the first type data is extracted based on a feature extraction large language model; second type data corresponding to the target power spot transaction volume prediction type is determined, and the second type data is vectorized to obtain a second feature vector based on a preset feature vectorization method, the second type data including at least one of the following: historical residential electricity consumption index, historical industrial electricity consumption index, electricity demand index, temperature and humidity index; a pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type is determined; according to the first feature vector and the second feature vector, and based on the pre-trained power spot transaction volume probability prediction model, the power spot transaction volume is predicted to obtain a power spot transaction volume prediction interval corresponding to the target power spot transaction volume prediction type.

2. The deep learning-based power spot transaction volume dynamic prediction method of claim 1, wherein, The method further comprises: in the absence of corresponding first type data, the power spot transaction volume is predicted according to the second feature vector and based on the pre-trained power spot transaction volume probability prediction model to obtain a power spot transaction volume prediction interval corresponding to the target power spot transaction volume prediction type. 3.The deep learning-based power spot trading volume dynamic prediction method according to claim 1 or 2, characterized in that, According to the target power spot transaction volume prediction type and based on a preset data identifier, it is determined whether there is corresponding first type data, including: determining a pre-trained power spot transaction volume probability prediction model corresponding to the target power spot transaction volume prediction type; determining whether the pre-trained power spot transaction volume probability prediction model is set with a preset data identifier; in the case where the preset data identifier is not set, it is determined that there is no corresponding first type data; in the case where the preset data identifier is set, it is determined whether there is target specific content data corresponding to the preset data identifier; in the absence of the target specific content data, it is determined that there is no corresponding first type data; in the presence of the target specific content data, it is determined that there is corresponding first type data.

4. The deep learning-based power spot transaction volume dynamic prediction method of claim 3, wherein, Determining whether the pre-trained power spot transaction volume probability prediction model is set with a preset data identifier includes: according to the pre-trained power spot transaction volume probability prediction model, it is determined whether a preset data switch corresponding to the preset data identifier is turned on; in the case where the preset data switch is turned on, it is determined that the pre-trained power spot transaction volume probability prediction model is set with a preset data identifier; in the case where the preset data switch is not turned on, it is determined that the pre-trained power spot transaction volume probability prediction model is not set with a preset data identifier.

5. The deep learning-based power spot trading volume dynamic prediction method of claim 4, wherein, The method further comprises: determining whether the preset data switch corresponding to the preset data identifier is opened, including: in response to the start of the dynamic prediction of the power spot transaction volume by the pre-trained power spot transaction volume probability prediction model, prompting a user to open the preset data switch based on a preset prompt page; detecting a configuration operation of the user on the preset data switch; in response to the configuration operation of the user opening the preset data switch, opening the preset data switch; and in response to the configuration operation of the user not opening the preset data switch, not opening the preset data switch.

6. The deep learning-based power spot trading volume dynamic prediction method of claim 3, wherein, Before determining whether the target specific content data corresponding to the preset data identifier exists, the method further comprises: monitoring whether the network information corresponding to the first type of data exists based on a preset legal and compliant network information crawling method; and in the case where the network information exists, obtaining the network information to obtain the target specific content data.

7. The deep learning-based power spot transaction volume dynamic prediction method according to claim 1 or 2, characterized in that, The method further comprises: determining the second type of data corresponding to the target power spot transaction volume prediction type, including at least one of the following: determining preset structured dynamic data corresponding to the target power spot transaction volume prediction type, the preset structured dynamic data including at least one of the following: the latest power spot transaction clearing result, the latest weather temperature and humidity; determining preset structured static data corresponding to the target power spot transaction volume prediction type, the preset structured static data including at least one of the following: historical power spot price, historical power spot transaction volume; determining preset unstructured static data corresponding to the target power spot transaction volume prediction type, the preset unstructured static data including historical policy and market rule change information; and accordingly, at least one of the preset structured dynamic data, the preset structured static data and the preset unstructured static data is taken as the second type of data. The method further comprises: performing vectorization on the second type of data based on a preset feature vectorization method to obtain a second feature vector, including: performing corresponding semantic information feature extraction based on a feature extraction large language model according to the preset unstructured static data to obtain a corresponding second feature vector.

8. The deep learning-based power spot transaction volume dynamic prediction method of claim 7, wherein, The method further comprises: predicting the power spot transaction volume based on the first feature vector and the second feature vector and the pre-trained power spot transaction volume probability prediction model to obtain a power spot transaction volume prediction interval corresponding to the target power spot transaction volume prediction type, including: predicting the power spot transaction volume based on the first feature vector and the second feature vector and the pre-trained power spot transaction volume probability prediction model to obtain an initial power spot transaction volume prediction interval and an explanatory feature vector corresponding to the initial power spot transaction volume prediction interval; determining a preset prompt word template corresponding to the pre-trained power spot transaction volume probability prediction model; embedding the initial power spot transaction volume prediction interval and the explanatory feature vector into the preset prompt word template to obtain a target prompt word; and generating a power spot transaction volume prediction interval corresponding to the initial power spot transaction volume prediction interval and an explanatory text based on the target prompt word and a text generation large language model. 9.The deep learning based power spot trading volume dynamic prediction method of claim 1 or 2, wherein, ​ 10. The deep learning-based power spot transaction volume dynamic prediction method of claim 9, wherein, The initial electricity spot transaction volume prediction interval and the explanatory feature vector are embedded into the preset prompt word template to obtain a target prompt word, including: generating a target query condition according to the initial electricity spot transaction volume prediction interval and the explanatory feature vector, and based on a preset query condition generation mode; performing retrieval enhancement on the target query condition to obtain target retrieval enhancement data according to the target query condition and based on a preset retrieval enhancement generation paradigm; and embedding the initial electricity spot transaction volume prediction interval, the explanatory feature vector and the target retrieval enhancement data into the preset prompt word template to obtain the target prompt word.

Images