Energy storage carbon emission reduction prediction method and device based on large language model

Abstract

The invention discloses an energy storage carbon emission reduction prediction method and device based on a large language model. The method is applied to an energy storage battery and comprises the steps that to-be-used data generated in the energy storage battery in the emission duration is obtained, and the data at least comprises one or more of operation and measurement type numerical data, unit and emission related structural data, market and carbon policy type text data and external environment load type data; determining a target feature tensor to be input into the target language model according to the to-be-used data, inputting the target feature tensor into the pre-trained target language model, and outputting the carbon emission and confidence attribute of the to-be-used data within the prediction duration; the target feature tensor at least comprises one or more of the charging power, the discharging power, the state of charge, the target marginal emission factor, the electric quantity transaction attribute, the environment feature and the electric energy policy file at the discrete time point. The energy storage carbon emission reduction amount is predicted based on the large language model, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage system technology, and in particular to a method and device for predicting carbon emission reductions from energy storage based on a large language model. Background Technology

[0002] As a core component for achieving smooth grid connection of renewable energy and peak shaving and valley filling in the power grid, the carbon emission level of energy storage batteries throughout their entire life cycle directly affects the emission reduction effectiveness of the power system. On the one hand, the large-scale grid connection of intermittent renewable energy sources such as wind and solar has led to a surge in demand for energy storage batteries. However, carbon emissions are present throughout the entire process of energy storage batteries, from mining and material preparation to manufacturing, transportation, operation and maintenance, and decommissioning and recycling. Ignoring their own carbon footprint could result in "high-carbon energy storage" offsetting the emission reduction benefits of renewable energy. On the other hand, energy storage batteries using different technologies (such as lithium-ion batteries, flow batteries, and lead-acid batteries) differ significantly in raw material types, production processes, cycle life, and recycling efficiency, resulting in drastically different carbon emission intensities. This necessitates the use of scientific carbon emission prediction technologies to quantify the carbon emissions at each stage. Current methods for predicting carbon emissions from energy storage only use annual average emission factors for static carbon emission prediction, resulting in poor accuracy. Furthermore, many methods rely on single models and ignore semantic documents, leading to poor accuracy and interpretability. Summary of the Invention

[0003] This invention provides a method and apparatus for predicting carbon emission reductions from energy storage based on a large language model, in order to solve the problems of poor accuracy and poor interpretability in predicting carbon emissions from energy storage.

[0004] According to one aspect of the present invention, a method for predicting carbon emission reduction in energy storage based on a large language model is provided, applicable to energy storage batteries, comprising:

[0005] Acquire the unused data generated in the energy storage battery during the emission period, wherein the unused data includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load;

[0006] Based on the data to be used, determine the target feature tensor to be input into the target language model, and input the target feature tensor into the pre-trained target language model to output the carbon emissions and confidence attributes of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

[0007] Optionally, the target marginal emission factor is determined based on the following methods: determining the change in state of charge (SOC) based on the target and initial stored energy from the operational and measurement data; correcting the rated capacity of the energy storage battery based on a preset correction coefficient when the design parameters from the operational and measurement data determine that the battery does not meet preset conditions, thus obtaining the usable energy storage capacity; determining the charging energy value based on the instantaneous charging power from the operational and measurement data, and determining the discharging energy value based on the instantaneous discharging power from the operational and measurement data; determining the charging efficiency and discharging efficiency based on the SOC, rated energy storage capacity, charging energy value, and discharging energy value; and determining the target marginal emission factor based on the charging efficiency, discharging efficiency, preset correction information, and the collected emission factors from the operational and measurement data.

[0008] Optionally, based on the data to be used, a target feature tensor to be input into the target language model is determined, and the target feature tensor is input into the pre-trained target language model to output the carbon emissions and confidence attributes of the data to be used within the prediction time period. This includes: determining the target feature tensor based on the data to be used corresponding to each discrete time point; performing feature processing on the target feature tensor based on the embedding layer in the target language model, and performing positional encoding on the feature vector after feature processing to obtain the feature vector to be input into the encoding layer in the target language model; after sequentially processing the feature vector based on at least one encoding layer, performing multi-head self-attention calculation on the obtained encoded features to obtain the attention-processed feature vector; after processing the attention-processed feature vector based on the residual connection and normalization layer in the target language model, inputting it into the feedforward network in the target language model to obtain the feature encoding sequence; and processing the feature encoding sequence based on the pooling layer in the target language model to obtain the carbon emissions and confidence attributes within the prediction time period.

[0009] Optionally, the method further includes: repeatedly executing the carbon emissions and confidence attribute preset number of times within the prediction period; determining the mean and variance of carbon emissions based on the carbon emissions corresponding to each time, and determining the confidence interval based on the mean and variance of carbon emissions; if the mean or variance of the confidence attribute and the confidence interval satisfy a preset relationship, then the prediction result is determined to meet the preliminary conditions.

[0010] Optionally, the method also includes: processing the target payoff function using core decision variables and charge / discharge constraints to determine the charge / discharge power sequence within the prediction period.

[0011] Optionally, the target payoff function is processed using core decision variables and charge / discharge constraints to determine the charge / discharge power sequence within the prediction period, including:

[0012] The first language model is used to process market and carbon policy text data in the data to be used, resulting in carbon benefit weights. These weights include carbon benefit information corresponding to different time periods. A particle swarm optimization algorithm is used, along with core decision variables and charging / discharging constraints, to process the target benefit function, yielding multiple candidate power sequences. The target benefit function must include at least carbon benefit information. Based on an iterative update of multiple candidate power sequences, the maximum value of the target benefit function is determined, and the candidate power sequence corresponding to this maximum value is used as the charging / discharging power sequence for the prediction period.

[0013] Optionally, the method further includes: generating an interpretable report for a preset object based on the charge and discharge power sequence within the predicted duration; wherein the interpretable report includes at least an overall revenue overview, time-segmented scheduling suggestions, risk warnings, and generation link information for generating the interpretable report, and the generation link information includes at least the normative documents on which it is based.

[0014] According to another aspect of the present invention, an energy storage carbon emission reduction prediction device based on a large language model is provided, comprising:

[0015] The data to be used module is used to acquire the data to be used generated in the energy storage battery during the emission period. The data to be used includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emission, textual data of market and carbon policy, and data of external environmental load.

[0016] The carbon emission and confidence attribute prediction module is used to determine the target feature tensor to be input into the target language model based on the data to be used, and input the target feature tensor into the pre-trained target language model to output the carbon emission and confidence attribute of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

[0017] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0018] At least one processor; and

[0019] A memory that is communicatively connected to at least one processor; wherein,

[0020] The memory stores a computer program that can be executed by at least one processor, such that the at least one processor is able to execute the energy storage carbon emission reduction prediction method based on a large language model according to any embodiment of the present invention.

[0021] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for causing a processor to execute and implement the energy storage carbon emission reduction prediction method based on a large language model according to any embodiment of the present invention.

[0022] The technical solution of this invention involves acquiring unused data generated in an energy storage battery during the emission period. This unused data includes at least one or more of the following: numerical data related to operation and measurement, structured data related to the unit and emissions, textual data related to the market and carbon policies, and data related to external environmental loads. Based on the unused data, a target feature tensor is determined to be input into a target language model. This target feature tensor is then input into a pre-trained target language model, outputting the carbon emissions and confidence attributes of the unused data within the prediction period. The target feature tensor includes at least one or more of the following: charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and data from electricity policy documents at discrete time points. This solution processes one or more of the following data over the emission period: numerical data related to operation and measurement, structured data related to units and emissions, textual data related to the market and carbon policies, and data related to external environmental loads. This process yields a target feature tensor, which is then adapted to the input data requirements of the target language model. This provides comprehensive and accurate foundational support for carbon emission prediction. Furthermore, leveraging the multi-feature analysis capabilities of a mature target language model, it enables collaborative analysis of numerical and textual information. Combined with the output of confidence attributes, this not only improves the accuracy and reliability of carbon emission prediction but also provides a scientific and effective decision-making basis for carbon emission control and policy compliance assessment of energy storage systems.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a method for predicting carbon emission reductions from energy storage based on a large language model, provided in Embodiment 1 of the present invention.

[0026] Figure 2 This is a flowchart of a method for predicting carbon emission reductions from energy storage based on a large language model, provided in Embodiment 2 of the present invention.

[0027] Figure 3 This is a schematic diagram of the structure of an energy storage carbon emission reduction prediction device based on a large language model provided in Embodiment 3 of the present invention;

[0028] Figure 4 This is a schematic diagram of the structure of an electronic device that implements the energy storage carbon emission reduction prediction method based on a large language model according to an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] Example 1

[0032] Figure 1 This is a flowchart of a method for predicting carbon emission reduction in energy storage based on a large language model, provided in Embodiment 1 of the present invention. This embodiment is applicable to situations involving carbon emission reduction prediction in energy storage. The method can be executed by an energy storage carbon emission reduction prediction device based on a large language model. This device can be implemented in hardware and / or software and can be configured in electronic devices such as computers, servers, and controllers. The method is applied to energy storage batteries, such as... Figure 1 As shown, the method includes:

[0033] S110. Obtain the unused data generated in the energy storage battery during the emission period, wherein the unused data includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load.

[0034] Specifically, energy storage batteries can be understood as the core energy storage component of an energy storage system, used to store and release electrical energy, and are widely used in scenarios such as grid peak shaving, new energy support, and emergency power supply. It is understandable that energy storage batteries will generate carbon emissions that pollute the environment throughout their entire life cycle; therefore, it is necessary to predict the carbon emissions of energy storage batteries. The emission duration can be understood as the time interval set for collecting the energy storage battery usage data needed to predict the carbon emissions of energy storage. To avoid the problem of collecting too much or too little data, the emission duration must be determined based on the energy storage battery's operating conditions and the prediction duration. It should be noted that, using the current moment as a benchmark, historical periods similar to the current operating conditions of the energy storage battery should be prioritized within the basic span. For example, if the current period is during the summer high-temperature period and the grid peak load regulation period, it is necessary to ensure that the emission duration includes high-temperature and peak load operating data from the same period in previous years. If the current battery is in a specific lifespan stage (e.g., 1000 cycles), historical data from newly commissioned batteries or those in the later stages of aging should be excluded to avoid prediction errors due to mismatched operating conditions. Alternatively, the predicted future timescale, i.e., the prediction duration, can be clearly defined first. The corresponding emission duration can then be set based on the prediction duration. For example, if the prediction is for the next 24 hours (short-term), the emission duration should ideally be based on historical data from the past 7–14 days. This span covers daily peak and off-peak charging / discharging times, as well as weekday / weekend operating conditions, without introducing excessive redundancy. If the prediction is for the next week (medium-term), the emission duration needs to be extended to the past 1–2 months to cover influencing factors such as weekly policy adjustments and energy price fluctuations. If the prediction is for more than one month (long-term), the emission duration should be based on historical usage data from the past 3–6 months to ensure coverage of seasonal fluctuations in carbon emissions, such as the impact of summer-winter temperature differences on battery efficiency. Optionally, given the determined emission durations corresponding to the energy storage battery's operating conditions and the prediction duration, a longer emission duration can be selected as the final emission duration. In this embodiment, the emission duration can be predetermined and stored in a designated storage space, and can be directly retrieved when data is needed.

[0035] Specifically, the system acquires usable data generated by the energy storage battery during the emission period through a pre-defined data access layer. This usable data includes at least one or more of the following: operational and measurement numerical data, unit-related and emission-related structured data, market and carbon policy text data, and external environmental load data. Operational and measurement numerical data includes, but is not limited to, energy storage data, battery charge / discharge power data, and design parameters. Design parameters include the comprehensive additional loss factor of the rectifier, inverter, and line losses, and the rated energy storage capacity. Unit-related and emission-related structured data includes, but is not limited to, unit model, capacity parameters, and unit emission factors. Market and carbon policy text data includes carbon market trading rules, carbon quota policies, electricity trading attribute adjustment notices, and unit maintenance plans. Environmental load data includes, but is not limited to, data on ambient temperature, humidity, and grid load changes in the energy storage battery's environment. Optionally, the acquired multi-source data is cleaned, deduplicated, formatted, and categorized to filter out valid usable data within the emission period, ultimately forming a structured dataset.

[0036] In this embodiment, multiple types of data, including equipment operation, unit emissions, policy and market, and external environment, are integrated, which can provide complete data support for the subsequent prediction of carbon emissions from energy storage batteries and help improve the accuracy of carbon emission prediction.

[0037] S120. Based on the data to be used, determine the target feature tensor to be input into the target language model, and input the target feature tensor into the pre-trained target language model, outputting the carbon emissions and confidence attributes of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

[0038] The target language model refers to a large language model pre-trained and optimized with a large amount of multi-source heterogeneous data. It possesses the ability to predict and analyze feature data that integrates numerical and textual information, and can specifically output prediction results related to carbon emissions in the energy storage field. The prediction results include carbon emissions and confidence attributes within the prediction period. The target feature tensor specifically refers to a multi-dimensional data structure formed by integrating key information from the data to be used, such as discrete-time point-of-use charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents, after standardization and vectorization preprocessing. It is the core data carrier adapted to the input requirements of the target language model. The prediction period refers to a pre-set future time interval used to predict the carbon emissions of energy storage batteries; its range can be determined based on actual needs and historical data patterns. It should be noted that electricity trading attributes can be the current electricity price data. Carbon emissions refer to the total carbon emissions, emissions per unit time, or carbon emission sequences generated by the energy storage system within the prediction period, predicted by the target language model based on the input feature tensor. Confidence attributes are indicators output by the model that characterize the reliability of the prediction results. They reflect the probability of the predicted value matching the actual value and can provide a reference for subsequent carbon emission control decisions.

[0039] Specifically, based on the previously collected data on emissions from energy storage batteries during the predicted period, key information such as charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points is selected. These heterogeneous data are then standardized, with numerical data normalized and textual data converted into vector form through word embedding. Subsequently, these vectors are integrated according to preset dimensions to construct a target feature tensor containing one or more key pieces of information. This target feature tensor is then input into a target language model trained on a large amount of historical emission data. The model uses its own algorithm architecture to perform deep analysis and computation on the feature tensor, ultimately outputting the carbon emission values ​​for the predicted period corresponding to the data to be used, as well as confidence attributes reflecting the reliability of the prediction results.

[0040] For example, based on the various data types in the preset data to be used, data from different business systems and different modalities (numerical and textual) can be aligned, cleaned, and characterized on a unified time axis and a unified space / unit dimension. Multi-source data within the emission period can be selected to provide a unified input for subsequent carbon emission reduction prediction and strategy optimization. The multi-source data specifically includes the following data: (1) Operation and measurement numerical data: power sequence of energy storage system (Positive indicates discharging, negative indicates charging); State variables such as energy storage SOC, battery temperature, and cycle count; Grid marginal emission factor. Day-ahead / real-time electricity prices and ancillary service prices (2) Structured data related to units and emissions: unit list, unit type (coal-fired, gas-fired, pumped storage, etc.), capacity, start-up and shutdown status; unit baseline emission factor, thermal efficiency and other operating parameters; marginal unit identity markers of different units in each time period (determining MEF). (3) Market and carbon policy text data (PDF / announcement / HTML): carbon emission trading rules, quota allocation and assessment methods; electricity market rules (e.g. peak-valley electricity price policy, ancillary service compensation rules); energy storage operation constraint documents, such as "no charging / no discharging" periods, market participation type restrictions; unit maintenance plans, temporary dispatch notices, etc. (4) External environment and load data (optional): historical load curves and holiday markers; meteorological data (temperature, irradiance, wind speed, etc.). Optionally, the above multi-source data are accessed to time series databases (e.g. TSDB) through different acquisition channels, and text data are jointly managed through file storage and vector databases. After obtaining multi-source data, in order to eliminate the inconsistencies in time scale, spatial / regional dimension and scale of different data sources, the present invention adopts the following fusion process: (1) Time alignment: Select a unified time granularity For example, resample each data source to a unified timeline over 15 minutes or 1 hour. For missing time periods, forward filling or interpolation is used. (2) Spatial / unit dimension alignment: The unit emission factor and start-up / shutdown status are aligned with the grid marginal emission factor by using unit ID, area code, voltage level and other keys. (3) Numerical normalization and unit unification: Unit conversion and normalization are performed on numerical features such as power, electricity trading attributes, and emission factors, for example, z-score and min-max algorithms, to improve the training stability of time series models. (4) Text to structured features: Key information is extracted from texts such as policy PDFs and maintenance announcements using a large language model + regular expressions. Key information includes: time intervals for charging / discharging bans. Carbon price preferences or policy inclinations correspond to carbon yield weights. The value range; changes in the "effective unit set" due to unit maintenance; temporary correction rules for emission factors or on-grid electricity prices of a certain type of unit. The extracted structured results are aligned with the time-series data by timestamp, serving as an additional feature or constraint set. The fused result can be abstracted into a unified feature tensor and constraint set:

[0041] ;

[0042] in, For the target feature tensor, The power at time t, such as the charging / discharging power of energy storage; Let be the state of charge at time t; The characteristics of the power grid at time t, such as the marginal emission factor of the power grid; Let be the electricity price at time t; For environmental characteristics at time t, such as temperature and humidity; The policy characteristics at time t; This represents the set of operational constraints that take effect at time t, such as prohibiting discharge / charging / discharge during the current time period, and adjusting the power limit, etc. It is a global constraint set library.

[0043] The obtained target feature tensor is input into the pre-trained target language model, which outputs the carbon emissions and confidence attributes of the data to be used within the prediction time period. Specifically, the current prediction time can be set as t, and a historical time window of length L=24 {t-L+1, ..., t} can be selected. The feature tensor at each time point is defined as follows:

[0044] ;

[0045] Let the feature sequence of length L be denoted as:

[0046] ;

[0047] The model output is a predicted sequence of emission reductions for the next H time steps (i.e., the prediction duration). and its confidence properties :

[0048] .

[0049] In this embodiment, by integrating multi-dimensional heterogeneous data into a feature tensor and inputting it into a well-trained language model, the fusion analysis of different types of data such as numerical and textual data is realized, improving the comprehensiveness and accuracy of carbon emission prediction. Furthermore, the reliability of the prediction results is clarified by the output of confidence attributes, providing a decision-making basis with both accuracy and reference value for subsequent work such as carbon emission control and policy compliance assessment of energy storage systems.

[0050] Optionally, the target marginal emission factor is determined based on the following methods: determining the change in state of charge (SOC) based on the target and initial stored energy from the operational and measurement data; correcting the rated capacity of the energy storage battery based on a preset correction coefficient when the design parameters from the operational and measurement data determine that the battery does not meet preset conditions, thus obtaining the usable energy storage capacity; determining the charging energy value based on the instantaneous charging power from the operational and measurement data, and determining the discharging energy value based on the instantaneous discharging power from the operational and measurement data; determining the charging efficiency and discharging efficiency based on the SOC, rated energy storage capacity, charging energy value, and discharging energy value; and determining the target marginal emission factor based on the charging efficiency, discharging efficiency, preset correction information, and the collected emission factors from the operational and measurement data.

[0051] Specifically, the change in state of charge (SOC) is calculated based on the target stored energy and initial stored energy in the operational and measurement numerical data. The calculation formula is as follows:

[0052] ;

[0053] in, express The stored energy at any given moment, as the initial stored energy. express Energy storage at any given moment, as the target energy storage.

[0054] Next, based on the design parameters in the measured numerical data, it is determined whether the energy storage battery meets the preset conditions. If not, a preset correction coefficient is used to correct the rated energy storage capacity to obtain the corresponding usable energy storage capacity. Specifically, it is determined whether the rated energy storage capacity in the design parameters meets the preset energy storage capacity threshold. If it does, the energy storage battery is determined to meet the preset conditions; otherwise, it is determined that the energy storage battery does not meet the preset conditions. If the energy storage battery meets the preset conditions, the preset correction coefficient is obtained, and the rated energy storage capacity is corrected using the preset correction coefficient to obtain the usable energy storage capacity. The expression corresponding to the correction is shown below:

[0055] ;

[0056] in, This indicates the rated capacity of the energy storage battery, and SOH represents the state of health coefficient of the energy storage battery. This indicates the available energy storage capacity.

[0057] Subsequently, the charging energy value and discharging energy value were calculated by combining the instantaneous charging power and instantaneous discharging power from the operational and measurement numerical data; the charging energy value... The calculation formula is as follows:

[0058] ,or, ;

[0059] in, The instantaneous charging power is continuously sampled. The instantaneous charging power at the k-th sampling time of the discrete sampling is... , This represents the start and end times of the charging time interval.

[0060] Discharge energy value The calculation formula is as follows:

[0061] ,or, ;

[0062] in, The instantaneous discharge power is continuously sampled. Let K be the instantaneous discharge power at the k-th sampling time of the discrete sampling. , These represent the start and end times of the discharge time interval.

[0063] Furthermore, the charging efficiency and discharging efficiency are derived based on the change in state of charge, rated energy storage capacity, charging energy value, and discharging energy value; among which, the charging efficiency... and discharge efficiency The calculation formula is as follows:

[0064] ;

[0065] ;

[0066] Finally, by integrating charging efficiency, discharging efficiency, preset correction information, and collected emission factors, the target marginal emission factor is calculated and determined. The calculation formula is as follows:

[0067] ;

[0068] in, Indicates the initial grid emission factor. This represents the baseline cycle efficiency, which can be 0.92. Lower real-time efficiency results in lower efficiency. The higher the value, the higher the input energy required to release 1 kWh of energy storage, thus improving the accuracy of carbon accounting.

[0069] In this embodiment, by combining the actual operating parameters of the energy storage battery with the capacity correction mechanism, the calculation of key indicators such as state of charge and rated energy storage capacity is made more consistent with the actual operating conditions of the battery. At the same time, the charging and discharging efficiency is incorporated into the factor calculation system, which realizes the dynamic adjustment of emission factors according to the actual state of the battery. This effectively improves the accuracy and rationality of the target marginal emission factors and provides more reliable basic parameter support for subsequent carbon emission prediction.

[0070] Based on the above embodiments, the method further includes: processing the target profit function using core decision variables and charging / discharging constraints to determine the charging / discharging power sequence within the prediction period.

[0071] In this context, core decision variables refer to key controllable parameters pre-selected or adjusted during the optimization of the energy storage system's charging and discharging strategy. These are the core controllable objects for achieving optimized returns and safe operation. In this embodiment, core decision variables include the charging and discharging power sequence of the energy storage battery at the predicted future time. Charging and discharging constraints can be specifically understood as various limiting conditions to ensure the stable and compliant operation of the energy storage battery, including but not limited to upper and lower limits of charging and discharging power, state of charge range, and depth of charge / discharge limits. These are boundary conditions that must be followed during the optimization calculation process. The target return function can be specifically understood as a mathematical expression used to quantify the comprehensive return of the energy storage battery within the predicted timeframe. The expression is as follows:

[0072] ;

[0073] in, Indicates the gain of electrical energy. Indicates carbon gain. This is the weighting factor for carbon revenue, used to adjust the proportion of carbon revenue in total revenue. It represents the variance of returns, signifying the risk of return volatility. This is the risk weight; the larger the variance, the greater the penalty. The total return is calculated using this function, and the ultimate goal is to maximize this total return. Therefore, the function value is maximized or optimized through optimization. , The calculation formula is as follows:

[0074] ;

[0075] ;

[0076] in, This refers to the electricity price for the corresponding t+k period. It is the energy storage charging and discharging power of the corresponding t+k period. The sum of the products of the two is the revenue from electricity trading. It is the product of the corrected carbon emission factor and the energy storage charging and discharging power, corresponding to the benefits in scenarios such as carbon trading.

[0077] The prediction duration can be understood as a pre-defined time interval used to formulate charging and discharging strategies and calculate returns. Its length needs to be determined in conjunction with actual scenarios such as market cycles and grid dispatch requirements. The charging and discharging power sequence represents the combination of charging and discharging power values ​​at different time points within the prediction duration, obtained by optimizing the objective return function under the premise of satisfying charging and discharging constraints. It is the core basis for guiding the energy storage system to perform charging and discharging operations sequentially, and the charging and discharging power sequence can be determined by improving particle swarm optimization and greedy insertion strategies.

[0078] Specifically, the core decision variables and charging / discharging constraints for the operation of the energy storage system are clearly defined. Then, a target return function is constructed that includes dimensions such as carbon trading revenue, electricity market arbitrage revenue, and risk costs. Subsequently, the core decision variables and charging / discharging constraints are substituted into the target return function. The target return function is solved and optimized through linear programming, nonlinear programming, or intelligent optimization algorithms (such as particle swarm optimization and genetic algorithms). Finally, the charging / discharging power sequence that balances revenue maximization and operational safety within the prediction time is determined.

[0079] In this embodiment, by incorporating the core decision variables and constraints into the optimization solution process of the benefit function, it is ensured that the charging and discharging power sequence meets the physical operating limits of the energy storage battery and the grid dispatch requirements, while also achieving the optimal economic benefit within the prediction period. This also improves the scientificity and feasibility of the energy storage system operation strategy.

[0080] Optionally, the target payoff function is processed using core decision variables and charging / discharging constraints to determine the charging / discharging power sequence within the prediction period. This includes: processing market and carbon policy text data in the data to be used based on a first-language model to obtain carbon payoff weights; wherein the carbon payoff weights include carbon payoff information corresponding to different time periods; processing the target payoff function using a particle swarm optimization algorithm, core decision variables, and charging / discharging constraints to obtain multiple candidate power sequences; wherein the target payoff function includes at least carbon payoff information; determining the maximum value of the target payoff function based on an iterative update of multiple candidate power sequences, and using the candidate power sequence corresponding to the maximum value as the charging / discharging power sequence within the prediction period.

[0081] Specifically, based on a first-language model, the market and carbon policy text data in the collected data to be used are parsed, semantically extracted, and value quantified to generate carbon benefit weights corresponding to different time periods and incorporate them into the target benefit function. Then, taking core decision variables such as charging and discharging power and start-stop time as optimization objects, and combined with charging and discharging constraints such as upper and lower limits of battery charging and discharging power, state of charge range, and grid interaction restrictions, the particle swarm optimization algorithm is used to iteratively optimize the target benefit function containing carbon benefit information to obtain multiple candidate power sequences that meet the constraints. Subsequently, by continuously updating and iterating these candidate power sequences and comparing the target benefit function values ​​corresponding to different sequences, the candidate power sequence that can make the function value reach the maximum value is selected and determined as the charging and discharging power sequence within the prediction time.

[0082] In this embodiment, the carbon revenue value in policy texts is mined using a first language model, making revenue calculation more aligned with market and policy guidance. At the same time, the global optimization capability of the particle swarm optimization algorithm is used to maximize the target revenue while satisfying various constraints. Furthermore, the iterative comparison of multiple candidate sequences further ensures the scientificity and optimality of the charge and discharge power sequence, effectively improving the economy and compliance of the energy storage system operation.

[0083] For example, in obtaining the future H-step electricity price sequence Carbon emission reduction prediction sequence After determining its confidence properties, the energy storage scheduling power sequence is realized through an "improved particle swarm optimization-greedy hybrid algorithm". The optimization process considers electricity price arbitrage revenue, carbon revenue, prediction confidence constraints, and SOC physical constraints simultaneously, aiming to achieve a balance between maximizing market returns and minimizing operational risks. The specific steps are as follows:

[0084] 1. Optimization target: Energy storage scheduling power sequence The core decision variable for optimization is the charge and discharge power sequence of the energy storage system in the next H steps:

[0085] ;

[0086] It should be noted that, Indicates discharge. This indicates that the device is charging.

[0087] The constraints include:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] in, This is for gatherings during periods when charging / discharging is prohibited.

[0093] The goal is to maximize the total future revenue function:

[0094] ;

[0095] Therefore, the optimization target is the total return function that balances electricity price arbitrage, carbon revenue, and uncertainty risk. .

[0096] 2. Optimization of technical methods: Particle swarm optimization + greedy algorithm hybrid strategy

[0097] This embodiment employs a hybrid strategy of "improved particle swarm optimization (PSO) + target-sensitive greedy insertion," combining global search capabilities with local feasibility correction capabilities. The steps are as follows:

[0098] (1) Particle swarm optimization search for electricity price arbitrage skeleton

[0099] Particle Swarm Optimization (PSO) is used to search for the globally optimal electricity price arbitrage scheduling sequence, with the objective of maximizing... Each particle represents a candidate power sequence:

[0100] ;

[0101] Particle updates follow:

[0102]

[0103] in, For the individual's optimal, For the group optimality, the power and SOC are projected and corrected after the update to satisfy physical constraints and no-charge / no-discharge constraints.

[0104] The output of the particle swarm optimization algorithm is the "optimal electricity price arbitrage skeleton": ;

[0105] (2) Large Language Model (LLM) analysis of policy documents to determine carbon payoff weights The system reads carbon market policy documents (such as "expansion of peak-valley carbon price difference" and "carbon quota adjustment notice"), and LLM extracts the following key information: peak carbon price. Valley carbon price Carbon price spread This week / monthly carbon yield bias. LLM maps this to carbon yield weights. :

[0106] ;

[0107] Used to amplify or weaken the corresponding carbon gain item .

[0108] (3) Greedy insertion strategy enhances carbon gain and feasibility

[0109] Obtaining the framework for electricity price arbitrage and weight Subsequently, this invention employs a target-sensitive greedy adjustment strategy to perform fine-tuning during periods of high carbon price differentials:

[0110] like High and Larger discharge power increases the discharge power. :

[0111] ;

[0112] If the carbon factor is low, increase the charging power appropriately. :

[0113] ;

[0114] Perform SOC projection after each fine-tuning:

[0115] ;

[0116] If the time period for violating the no-charging / no-playing rule is adjusted, then it will remain unchanged. .

[0117] The essence of a greedy strategy is to Make minimal modifications to improve:

[0118] .

[0119] 3. Summary of the optimization process

[0120] The overall process is as follows:

[0121] The input data is: ;

[0122] Particle Swarm Optimization: Obtaining a scheduling framework that maximizes electricity price arbitrage profits .

[0123] LLM Policy Analysis: Obtaining Carbon Benefit Weights .

[0124] Target-sensitive greedy fine-tuning: Insert enhanced charge / discharge actions during periods of high carbon price difference to meet SOC and operating boundary conditions.

[0125] Generate the final scheduling sequence:

[0126] ;

[0127] in, This is a greedy fine-tuning term. The final output is... This is the optimal dispatch strategy that takes into account electricity price revenue, carbon revenue, and risk constraints, and can be directly applied to the actual operation and market application of energy storage equipment.

[0128] Based on the above embodiments, the method further includes: generating an interpretable report for a preset object based on the charge and discharge power sequence within the predicted time period; wherein the interpretable report includes at least an overall revenue overview, time-segmented scheduling suggestions, risk warnings, and generation link information for generating the interpretable report, and the generation link information includes at least the normative documents on which it is based.

[0129] The pre-defined target audience can be understood as the designated recipients of the interpretable report, including but not limited to energy storage system operators, grid dispatching departments, carbon trading management agencies, and other entities related to energy storage dispatching decisions. The overall revenue overview can be understood as a comprehensive summary of the combined revenue that the energy storage system can obtain by executing the charge and discharge power sequence within the forecast period, covering the total value and proportion distribution of various revenues such as electricity market arbitrage revenue, carbon trading revenue, and ancillary service revenue. The time-segmented dispatching recommendations can be understood as specific charging and discharging operation guidelines given for different time nodes within the forecast period based on the charge and discharge power sequence, including but not limited to the charging and discharging power magnitude, start-up and shutdown timing, and state-of-charge control range for each time period. The risk warnings can be understood as early warnings of various potential problems that may be encountered during the execution of the charge and discharge strategy, including equipment damage risks caused by overcharging and discharging of batteries, compliance risks caused by policy changes, and revenue fluctuation risks caused by market price fluctuations. The generation link information can be understood as a complete traceability description of the process of forming an interpretable report, including but not limited to the source of the basic data on which the report is based, the algorithm model, the optimization logic, and the normative documents such as power dispatching specifications and carbon emission reduction policies on which the report is based, in order to ensure the scientific and compliant nature of the report.

[0130] Specifically, based on the optimal charging and discharging power sequence determined within the predicted time period, and combined with the calculation results of the objective revenue function, an overall revenue overview is summarized. Then, based on the power allocation, carbon revenue weight, and market policy requirements for different time periods, time-sharing scheduling suggestions are extracted. At the same time, the potential battery loss risk and policy compliance risk caused by power adjustment are analyzed and risk warnings are given. Subsequently, the core data sources, algorithm models, optimization logic, and normative documents such as electricity market trading rules and carbon emission reduction policies on which the entire process is based are traced and integrated to form generation link information. Finally, the overall revenue overview, time-sharing scheduling suggestions, risk warnings, and generation link information are systematically arranged to generate an interpretable report for the preset target audience.

[0131] In this embodiment, by integrating multi-dimensional key information to form a standardized report, the benefits and implementation plan of the charging and discharging strategy are clearly presented. The transparency and compliance of the strategy are enhanced by generating link information. At the same time, risk warnings can help the target to avoid potential problems in advance, and improve the scientificity and feasibility of energy storage system scheduling decisions.

[0132] The technical solution of this embodiment relies on multiple sources of terminals, such as the battery management system of the energy storage battery, emission monitoring equipment, policy information platform, and environmental sensor network, to collect one or more of the following data within the emission period: numerical data of operation and measurement, structured data related to units and emissions, textual data of the market and carbon policies, and data of external environmental load, to form data to be used. Then, key information such as charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points are extracted from the data to be used. After preprocessing such as numerical normalization and text vectorization, the data is integrated to construct a target feature tensor, which is input into a pre-trained target language model. The model performs in-depth analysis and calculation on the multi-dimensional features, and finally outputs the carbon emissions within the prediction period and the confidence attributes representing the reliability of the results. This solution processes one or more of the following data over the emission period: numerical data related to operation and measurement, structured data related to units and emissions, textual data related to the market and carbon policies, and data related to external environmental loads. This process yields a target feature tensor, which is then adapted to the input data requirements of the target language model. This provides comprehensive and accurate foundational support for carbon emission prediction. Furthermore, leveraging the multi-feature analysis capabilities of a mature target language model, it enables collaborative analysis of numerical and textual information. Combined with the output of confidence attributes, this not only improves the accuracy and reliability of carbon emission prediction but also provides a scientific and effective decision-making basis for carbon emission control and policy compliance assessment of energy storage systems.

[0133] Example 2

[0134] Figure 2This is a flowchart of a method for predicting carbon emission reduction in energy storage based on a large language model, provided in Embodiment 2 of the present invention. The method in this embodiment is a further optimization of the method in the above embodiments. Optionally, based on the data to be used corresponding to each discrete time point, a target feature tensor is determined; feature processing is performed on the target feature tensor based on the embedding layer in the target language model, and the positional encoding is performed on the feature vector after feature processing to obtain the feature vector to be input into the encoding layer in the target language model; after sequential processing of the feature vector based on at least one encoding layer, multi-head self-attention calculation is performed on the obtained encoded features to obtain the attention-processed feature vector; after processing the attention-processed feature vector based on the residual connection and normalization layer in the target language model, it is input into the feedforward network in the target language model to obtain a feature encoding sequence; the feature encoding sequence is processed based on the pooling layer in the target language model to obtain the carbon emission amount and confidence attribute within the prediction time. Figure 2 As shown, the method includes:

[0135] S210. Obtain the unused data generated in the energy storage battery during the emission period, wherein the unused data includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load.

[0136] S220. Based on the data to be used corresponding to each discrete time point, determine the target feature tensor, wherein the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at the discrete time points.

[0137] In this context, discrete time points can be understood as specific, pre-defined, independent, and sequentially arranged time nodes during data collection and analysis of the energy storage battery's operation. The intervals between these nodes can be set according to actual needs, such as at the minute or hour level. No specific limitation is made here. Each discrete time point corresponds to a set of energy storage battery data to be used, serving as the fundamental time unit for constructing time-series feature tensors and achieving precise time-segmented analysis.

[0138] Specifically, the data to be used within the emission period of the energy storage battery is first split according to the time dimension. One or more of the following are located for each discrete time point: numerical data of operation and measurement, structured data related to unit and emission, textual data of market and carbon policy, and data of external environmental load. Then, key information such as charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents are extracted from the data of each discrete time point. The numerical data is normalized and the textual data is preprocessed by word embedding and vectorization. Finally, the processed data of each discrete time point is integrated according to the time dimension into a multi-dimensional target feature tensor that adapts to the input requirements of the target language model.

[0139] For example, let the current prediction time be t, and select a historical time window {t-L+1, ..., t} of length L=24. Define the feature tensor at each time as:

[0140] ;

[0141] Let the feature sequence of length L be denoted as:

[0142] ;

[0143] In this embodiment, data extraction and tensor construction are performed on a discrete time point basis. This ensures the temporal and accuracy of the feature tensor, enabling precise matching of the operating conditions and environmental policy conditions at each time point. Furthermore, by integrating multi-dimensional key information, the feature tensor becomes more representative, providing a high-quality data carrier for the accurate prediction of subsequent carbon emissions.

[0144] S230. Based on the embedding layer in the target language model, the target feature tensor is processed, and the processed feature vector is positionally encoded to obtain the feature vector to be input into the encoding layer of the target language model.

[0145] In the target language model, the embedding layer is responsible for mapping heterogeneous input features (target feature tensors) into low-dimensional dense vectors. Its core function is to transform discrete or heterogeneous data, which is difficult for the model to directly recognize, into feature vectors with semantic relevance and uniform dimension. This effectively reduces data dimensionality and computational complexity, laying the foundation for subsequent feature analysis. The encoding layer, on the other hand, is the core network layer responsible for deep feature extraction and temporal correlation analysis of the feature vectors output by the embedding layer, which contain positional encoding information. It is used to uncover potential patterns and correlations between features at different discrete time points, thus providing crucial feature support for the final prediction of carbon emissions and confidence attributes.

[0146] Specifically, the target feature tensor is first input into the embedding layer of the target language model. Through the mapping operation of the embedding layer, the heterogeneous numerical and textual features are transformed into low-dimensional dense feature vectors with a unified dimension and semantic association. Then, position encoding is performed on these feature vectors, and temporal position information is added to the vectors according to the order of each discrete time point, so as to clarify the temporal correlation of features at different time points. Finally, feature vectors that can be directly input into the model encoding layer are generated.

[0147] For example, first, the original features Embedded into a single linear mapping or a small feedforward network 3D feature space:

[0148] ;

[0149] And add learnable or fixed positional coding. The encoder input is obtained as follows:

[0150] .

[0151] In this embodiment, the processing of the embedding layer realizes the standardization and vectorization fusion of heterogeneous features, which solves the problem of difficult collaborative analysis of multiple types of data. The location encoding gives the feature vector temporal attributes, enabling the model to accurately capture the temporal patterns of energy storage operation data at different discrete time points, thereby improving the accuracy and reliability of subsequent carbon emission prediction.

[0152] S240. After sequentially processing the feature vector based on at least one coding layer, multi-head self-attention calculation is performed on the obtained coded features to obtain the attention-processed feature vector.

[0153] Multi-head self-attention can be understood as a computational mechanism in target language models that allows the model to simultaneously focus on the correlation of information at different positions and dimensions of the input feature vector. It sets up multiple parallel attention heads, each assigning weights to the encoded feature vector from different perspectives, quantifying the dependencies between features such as charging power, state of charge, and policy text vectors at different discrete time points. The calculation results from each attention head are then concatenated and linearly transformed to uncover the complex coupling patterns between features. The feature vector after attention processing is the feature carrier output after multi-head self-attention computation. While retaining the core information of the original feature vector, it incorporates the correlation weights and temporal logic between different features, highlighting key features that have a greater impact on carbon emission prediction and strengthening the temporal correlation of data at different discrete time points. It serves as an important basis for the model's subsequent output of carbon emission figures and confidence attributes.

[0154] Specifically, the feature vectors processed by the embedding layer and positional encoding are input into at least one encoding layer of the target language model. The neural network structure of the encoding layer completes the initial feature extraction and information integration, resulting in encoded features containing multi-dimensional temporal features of energy storage operation. Subsequently, multi-head self-attention computation is performed on these encoded features. That is, multiple parallel attention heads capture the correlation weights between different discrete time points and different dimensional features (such as charging power and carbon policy attributes) in the feature vector. The computation results of each attention head are then concatenated and linearly transformed to finally generate the attention-processed feature vector.

[0155] For example, a multi-head self-attention mechanism is set up, such as h=8. In the l-th layer encoder For the sequence Perform multi-head self-attention calculation. The k-th attention head... The calculation is as follows:

[0156] ;

[0157] ;

[0158] in, For the reason The matrix formed This is a single-head dimension. The outputs of h=8 heads are concatenated and linearly transformed to generate the attention-processed feature vector:

[0159] .

[0160] In this embodiment, the sequential processing of the encoding layer can mine the deep representation of features layer by layer, while the multi-head self-attention computation can simultaneously focus on the correlation information of different scales and dimensions in the feature vector, accurately capture the temporal patterns and multi-feature coupling relationships of energy storage operation data, and enable the model to more comprehensively understand the internal logic between data, thereby improving the accuracy of subsequent carbon emission prediction.

[0161] S250. After processing the feature vectors after attention processing based on the residual connections and normalization layers in the target language model, the vectors are input into the feedforward network in the target language model to obtain the feature encoding sequence.

[0162] The residual connection and normalization layer are key components in the target language model used to optimize feature propagation and improve training stability. The residual connection establishes a shortcut path across layers, directly adding the original feature vector (without complex calculations) to the feature vector after attention calculation, effectively mitigating the gradient vanishing problem as model depth increases and ensuring complete feature information transmission. The normalization layer follows the residual connection, standardizing the data distribution of the fused feature vector, eliminating the influence of excessive differences in feature values ​​across different dimensions, and accelerating model convergence. The feedforward network refers to the fully connected network structure in the target language model that performs deep information mining on the feature vector after residual connection and normalization. It typically consists of two linear transformation layers and one non-linear activation function layer, capable of performing complex mapping operations on the input feature vector to further extract deep correlations between features, ultimately outputting a feature encoding sequence with stronger representational capabilities, providing core support for subsequent carbon emission and confidence attribute predictions.

[0163] Specifically, the feature vectors obtained through multi-head self-attention computation are input into the residual connection and normalization layer of the target language model. The residual connection directly preserves the core information of the original feature vectors, avoiding the gradient vanishing problem during model training. At the same time, the normalization operation is used to standardize the distribution of the feature vectors, improving the stability of the data and the convergence speed of the model. Subsequently, the feature vectors optimized by the residual connection and normalization layer are input into the feedforward network of the model. Through the linear transformation and nonlinear activation function of the feedforward network, the features are further deeply mapped and information integrated, ultimately generating a feature encoding sequence that combines temporal correlation and feature coupling rules.

[0164] For example, residual connections and layer normalization are used:

[0165] ;

[0166] Feedforward networks are stacked in layers, with each layer also containing a position-wise feedforward network:

[0167] ;

[0168] in, This involves using two fully connected layers with nonlinear activation techniques, such as ReLU / GELU. After stacking multiple encoder layers, the final encoded sequence is obtained. .

[0169] In this embodiment, the combination of residual connections and normalization layers not only ensures the integrity of feature information but also improves the training efficiency and stability of the model. The deep processing of the feedforward network can further explore the deep correlation between features, making the final output feature encoding sequence more representative and providing high-quality feature support for the accurate prediction of subsequent carbon emissions and confidence attributes.

[0170] S260. Based on the pooling layer in the target language model, the feature encoding sequence is processed to obtain the carbon emissions and confidence attributes within the prediction time.

[0171] In the target language model, the pooling layer is a network layer that aggregates, compresses, and extracts key information from the feature encoding sequence output after processing by the feedforward network. Located at the end of the model feature extraction process, its core function is to reduce the dimensionality of sequence data containing multiple discrete time points and multi-dimensional features. Through methods such as global average pooling and global max pooling, redundant details in the feature encoding sequence are discarded, while core information that can characterize the overall operating conditions and feature correlation patterns of the energy storage system within the prediction time is retained. These core aggregated features are then passed to the model output layer to achieve the final prediction of the specific value and confidence attribute of carbon emissions. At the same time, it can also effectively reduce the computational complexity of the model and alleviate the overfitting problem.

[0172] Specifically, the feature encoding sequence obtained after residual connection, normalization and feedforward network processing is input into the pooling layer of the target language model. The pooling layer aggregates and compresses the multidimensional temporal features in the sequence through global average pooling or global max pooling, and extracts the core feature information that can characterize the overall operating conditions of energy storage within the prediction period. Then, combined with the mapping operation of the model output layer, the specific value of carbon emissions within the prediction period and the confidence attribute reflecting the reliability of the prediction result are finally output.

[0173] For example, in order to obtain the emission reduction prediction for the next H steps from the encoded sequence, this invention adopts a "window encoding + readout head" method, which uses the encoded vector of the last time step... As a compressed representation of the entire window; or pooling of the encoded vectors across all time steps (e.g., average pooling):

[0174] ;

[0175] Mapping the aggregated representation to future H-step emission reduction predictions:

[0176] ;

[0177] in, and These are learnable parameters.

[0178] Overall, it can be summarized as follows:

[0179] ;

[0180] in, Indicates having parameters The timing Transformer encoder and readout network.

[0181] In this embodiment, the pooling layer's aggregation processing can effectively reduce feature dimensionality and redundant information, while preserving key patterns in the feature encoding sequence, avoiding overfitting problems in the model, and making the output carbon emission prediction more accurate. Meanwhile, the synchronously output confidence attribute can provide a reliable reference for the carbon emission control decision of the energy storage system, improving the scientific nature of the decision.

[0182] Based on the above embodiments, the method further includes: repeatedly executing the carbon emission amount and confidence attribute preset number of times within the prediction period; determining the mean and variance of carbon emissions based on the carbon emission amount corresponding to each time, and determining the confidence interval based on the mean and variance of carbon emissions; if the mean or variance of the confidence attribute and the confidence interval satisfy a preset relationship, then the prediction result is determined to meet the preliminary conditions.

[0183] The "preset number of times" refers to the predetermined number of times the carbon emission prediction process, involving repeated execution of the carbon emission and confidence attribute prediction, will be performed to verify the robustness of the carbon emission prediction results. This number can be determined based on factors such as model accuracy requirements and computational resources, thereby obtaining multiple sets of prediction data for subsequent statistical analysis. The "preset relationship" refers to the pre-defined criteria used to determine the validity of the prediction results. Specifically, it refers to the matching rules that must be met between the mean or variance of the confidence attributes obtained from multiple predictions and the confidence interval of carbon emissions, such as the inclusion relationship of numerical intervals and the deviation threshold range. The "preliminary condition" refers to the criteria for determining the basic validity of the carbon emission prediction results after verification based on the preset relationship. Meeting this condition indicates that the error of the prediction results is within an acceptable range and can provide a reference for subsequent carbon emission management of energy storage systems; otherwise, the model needs to be optimized or the prediction process adjusted.

[0184] Specifically, the carbon emission and confidence attribute prediction operations are repeated a preset number of times according to the established process to obtain multiple sets of corresponding carbon emission values ​​and confidence attribute data. Then, the mean and variance of these carbon emission data are calculated based on statistical methods, and the confidence interval of carbon emissions is further determined by combining the mean and variance. Subsequently, the mean or variance of the confidence attributes obtained from multiple predictions is compared with the confidence interval. If the two meet the preset matching relationship, the carbon emission prediction result is determined to meet the preliminary validity conditions.

[0185] For example, in the prediction (inference) phase, this embodiment also employs Monte Carlo Dropout technology to quantify prediction uncertainty, with the specific steps as follows:

[0186] During the inference phase, Dropout activation is maintained, meaning that some neuron connections are randomly dropped during each forward propagation; this applies to the same input window. Perform N independent forward propagations (N=30 in this embodiment) to obtain a set of prediction sequences:

[0187] ;

[0188] For each future time step Calculate the predicted mean and variance:

[0189] ;

[0190] ;

[0191] Construction based on variance Confidence interval:

[0192] ;

[0193] Therefore, this embodiment not only obtains point predictions of future emission reductions. It also obtains the variance of the prediction at each time step. confidence interval This allows for the quantification of prediction confidence attributes. Optionally, the format of the prediction results and the application scenario...

[0194] The final prediction results can be uniformly expressed as:

[0195] ;

[0196] These results can be applied in the system, including but not limited to: collaborative benefit optimization: point prediction. The model is used to calculate future carbon emission reduction benefits, which, together with electricity price arbitrage benefits, constitute the optimization objective. Risk constraints and conservative strategy generation: During periods with large variance, the optimization module introduces risk penalty terms or narrows the feasible scheduling interval to avoid adopting aggressive strategies under high uncertainty. Strategy interpretability and auditability: The confidence interval serves as a "prediction credibility" indicator, which can be explained in natural language by LLM, such as "The uncertainty of emission reduction prediction is high during a certain period; it is recommended to reduce the declared power," to meet market supervision and compliance audit requirements. Model adaptive updates: During periods of high uncertainty, model retraining or feature engineering adjustments can be triggered to continuously improve prediction performance. This embodiment's emission reduction prediction model, while ensuring high-precision point prediction, introduces quantifiable uncertainty estimation, significantly improving the robustness and interpretability of strategy generation.

[0197] In this embodiment, a statistical verification mechanism is introduced by repeatedly predicting the data, which effectively reduces the impact of random errors in a single prediction on the results. At the same time, by using multi-dimensional verification of the mean, variance, and confidence interval, the reliability and robustness of the prediction results are improved, providing more convincing data support for the formulation of subsequent carbon emission control strategies for energy storage systems.

[0198] The technical solution of this embodiment acquires the unused data generated by the energy storage battery during the emission period. This data includes one or more of the following: numerical data of operation and measurement, structured data related to the unit and emissions, textual data of the market and carbon policy, and data of external environmental load. Based on the unused data corresponding to each discrete time point, key information such as charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents are extracted and a target feature tensor is constructed. This tensor is then input into the target language model, first processed by the embedding layer and then positionally encoded to generate a feature vector that can be input into the encoding layer. The generated encoded features are then processed sequentially by at least one encoding layer, and multi-head self-attention calculation is performed on the generated encoded features. After optimization by the residual connection and normalization layer, the feature is fed into the feedforward network to obtain the feature encoding sequence. Finally, the pooling layer processes the sequence and outputs the carbon emissions during the prediction period and the confidence attributes of the results. This solution integrates various data types, including numerical data from operational measurements, structured data from unit emissions, textual data from market carbon policies, and data on external environmental loads. This comprehensive and multi-perspective approach captures key variables influencing carbon emissions, avoiding the limitations of single data sources. Furthermore, through core modules such as the embedding layer of the target language model, positional encoding, multi-head self-attention, residual connections, and normalization, it not only transforms heterogeneous feature tensors into feature vectors suitable for model analysis but also accurately captures potential correlations and temporal patterns between different features through multi-head self-attention. Residual connections and normalization effectively alleviate the gradient vanishing problem during model training, ensuring the stability and effectiveness of feature extraction and making feature encoding more accurate. Moreover, it not only outputs the carbon emissions within the predicted timeframe but also simultaneously provides confidence attributes representing the reliability of the results. This design provides clear data support for practical work such as carbon emission accounting and energy storage scheduling optimization, and helps relevant personnel determine the credibility range of the prediction results, providing a more rigorous reference for subsequent decisions. This improves the practicality and reliability of the prediction results and provides a more comprehensive reference for relevant decisions.

[0199] Example 3

[0200] Figure 3 This is a schematic diagram of the structure of an energy storage carbon emission reduction prediction device based on a large language model, provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes:

[0201] The data to be used acquisition module 310 is used to acquire the data to be used generated in the energy storage battery during the emission period. The data to be used includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emission, textual data of market and carbon policy, and data of external environmental load.

[0202] The carbon emission and confidence attribute prediction module 320 is used to determine the target feature tensor to be input into the target language model based on the data to be used, and input the target feature tensor into the pre-trained target language model, and output the carbon emission and confidence attribute of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

[0203] The technical solution of this embodiment acquires the data to be used generated in the energy storage battery during the emission period through the data acquisition module. The data to be used includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load. The carbon emission and confidence attribute prediction module determines the target feature tensor to be input into the target language model based on the data to be used, and inputs the target feature tensor into the pre-trained target language model to output the carbon emission and confidence attribute of the data to be used within the prediction period. The target feature tensor includes at least one or more of the following: charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points. This solution processes one or more of the following data over the emission period: numerical data related to operation and measurement, structured data related to units and emissions, textual data related to the market and carbon policies, and data related to external environmental loads. This process yields a target feature tensor, which is then adapted to the input data requirements of the target language model. This provides comprehensive and accurate foundational support for carbon emission prediction. Furthermore, leveraging the multi-feature analysis capabilities of a mature target language model, it enables collaborative analysis of numerical and textual information. Combined with the output of confidence attributes, this not only improves the accuracy and reliability of carbon emission prediction but also provides a scientific and effective decision-making basis for carbon emission control and policy compliance assessment of energy storage systems.

[0204] Based on the above embodiments, optionally, the carbon emission and confidence attribute prediction module 320 includes a target marginal emission factor determination unit. This unit is used to determine the change in state of charge (SOC) based on the target energy storage and initial energy storage in the operational and measurement-based numerical data; if the energy storage battery does not meet preset conditions based on the design parameters in the operational and measurement-based numerical data, it corrects the rated energy storage capacity of the energy storage battery based on a preset correction coefficient to obtain the usable energy storage capacity; it determines the charging energy value based on the instantaneous charging power in the operational and measurement-based numerical data, and determines the discharging energy value based on the instantaneous discharging power in the operational and measurement-based numerical data; it determines the charging efficiency and discharging efficiency based on the SOC, rated energy storage capacity, charging energy value, and discharging energy value; and it determines the target marginal emission factor based on the charging efficiency, discharging efficiency, preset correction information, and the collected emission factors in the operational and measurement-based numerical data.

[0205] Optionally, the carbon emission and confidence attribute prediction module 320 is specifically used to determine the target feature tensor based on the data to be used at each discrete time point; perform feature processing on the target feature tensor based on the embedding layer in the target language model, and perform position encoding on the feature vector after feature processing to obtain the feature vector to be input into the encoding layer in the target language model; after sequentially processing the feature vector based on at least one encoding layer, perform multi-head self-attention calculation on the obtained encoded features to obtain the attention-processed feature vector; after processing the attention-processed feature vector based on the residual connection and normalization layer in the target language model, input it into the feedforward network in the target language model to obtain the feature encoding sequence; and process the feature encoding sequence based on the pooling layer in the target language model to obtain the carbon emission and confidence attribute within the prediction time.

[0206] Optionally, the carbon emission and confidence attribute prediction module 320 is also used to repeatedly execute the carbon emission and confidence attribute preset number of times within a predetermined prediction period; determine the mean and variance of carbon emissions based on the carbon emission corresponding to each time, and determine the confidence interval based on the mean and variance of carbon emissions; if the mean or variance of the confidence attribute and the confidence interval satisfy a preset relationship, then the prediction result is determined to meet the preliminary conditions.

[0207] Optionally, the device is also used to process the target payoff function using core decision variables and charge / discharge constraints to determine the charge / discharge power sequence within the prediction period.

[0208] Optionally, the device is further configured to process market and carbon policy text data in the data to be used based on a first language model to obtain carbon benefit weights; wherein the carbon benefit weights include carbon benefit information corresponding to different time periods; using a particle swarm optimization algorithm, core decision variables, and charging / discharging constraints, the target benefit function is processed to obtain multiple candidate power sequences; wherein the target benefit function includes at least carbon benefit information; based on the iterative update of multiple candidate power sequences, the maximum value corresponding to the target benefit function is determined, and the candidate power sequence corresponding to the maximum value is used as the charging / discharging power sequence within the prediction time period.

[0209] Optionally, the device is also specifically used to generate an interpretable report for a preset object based on the charge and discharge power sequence within the predicted duration; wherein the interpretable report includes at least an overall revenue overview, time-segmented scheduling suggestions, risk warnings, and generation link information for generating the interpretable report, and the generation link information includes at least the normative documents on which it is based.

[0210] The energy storage carbon emission reduction prediction device based on a large language model provided in this embodiment of the invention can execute the energy storage carbon emission reduction prediction method based on a large language model provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0211] Example 4

[0212] Figure 4 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0213] like Figure 4As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0214] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0215] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the energy storage carbon emission reduction prediction method based on large language models.

[0216] In some embodiments, the large language model-based energy storage carbon emission reduction prediction method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the large language model-based energy storage carbon emission reduction prediction method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the large language model-based energy storage carbon emission reduction prediction method by any other suitable means (e.g., by means of firmware).

[0217] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0218] Computer programs for implementing the large language model-based carbon emission reduction prediction method for energy storage of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0219] Example 5

[0220] Embodiment 5 of the present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute a carbon emission reduction prediction method for energy storage based on a large language model, applicable to energy storage batteries. The method includes:

[0221] Acquire the unused data generated in the energy storage battery during the emission period, wherein the unused data includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load;

[0222] Based on the data to be used, determine the target feature tensor to be input into the target language model, and input the target feature tensor into the pre-trained target language model to output the carbon emissions and confidence attributes of the data to be used within the prediction period. The target feature tensor includes at least one or more of the following at discrete time points: charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and information from electricity policy documents.

[0223] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0224] To provide interaction with an object, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the object; and a keyboard and pointing device (e.g., a mouse or trackball) through which the object provides input to the electronic device. Other types of devices can also be used to provide interaction with the object; for example, feedback provided to the object can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the object can be received in any form (including sound input, voice input, or tactile input).

[0225] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., a computer with a graphical user interface or web browser through which an item can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0226] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0227] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0228] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting carbon emission reductions from energy storage based on a large language model, characterized in that, Applications in energy storage batteries include: Acquire the unused data generated in the energy storage battery during the emission period, wherein the unused data includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emissions, textual data of market and carbon policy, and data of external environmental load; Based on the data to be used, a target feature tensor to be input into the target language model is determined, and the target feature tensor is input into the pre-trained target language model to output the carbon emissions and confidence attributes of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

2. The method according to claim 1, characterized in that, The target marginal emission factor is determined based on the following method: The change in state of charge is determined based on the target energy storage and initial energy storage in the operational and measurement numerical data. If the design parameters in the operation and measurement numerical data determine that the energy storage battery does not meet the preset conditions, the rated energy storage capacity of the energy storage battery is corrected based on the preset correction coefficient to obtain the usable energy storage capacity. The charging energy value is determined based on the instantaneous charging power in the operation and measurement numerical data, and the discharge energy value is determined based on the instantaneous discharging power in the operation and measurement numerical data. The charging efficiency and discharging efficiency are determined based on the change in state of charge, the rated energy storage capacity, the charging energy value, and the discharging energy value. The target marginal emission factor is determined based on the charging efficiency, the discharging efficiency, the preset correction information, and the collected emission factors in the operation and measurement numerical data.

3. The method according to claim 1, characterized in that, The step of determining the target feature tensor to be input into the target language model based on the data to be used, inputting the target feature tensor into the pre-trained target language model, and outputting the carbon emissions and confidence attributes of the data to be used within the prediction time period includes: The target feature tensor is determined based on the data to be used at each discrete time point; Based on the embedding layer in the target language model, the target feature tensor is processed, and the processed feature vector is positionally encoded to obtain the feature vector to be input into the encoding layer of the target language model. After sequentially processing the feature vector based on at least one coding layer, multi-head self-attention calculation is performed on the obtained coded features to obtain the attention-processed feature vector. The feature vector after attention processing is processed by the residual connections and normalization layers in the target language model and then input into the feedforward network in the target language model to obtain the feature encoding sequence. The feature encoding sequence is processed based on the pooling layer in the target language model to obtain the carbon emissions and confidence attributes within the prediction time.

4. The method according to claim 3, characterized in that, The method further includes: Repeat the process of determining the carbon emissions and confidence attributes within the predicted time period a preset number of times; Based on the corresponding carbon emissions for each instance, determine the mean and variance of carbon emissions, and then determine the confidence interval based on the mean and variance of carbon emissions. If the mean or variance of the confidence attribute and the confidence interval satisfy a preset relationship, then the prediction result is determined to meet the preliminary conditions.

5. The method according to claim 1, characterized in that, The method further includes: The target reward function is processed using core decision variables and charge / discharge constraints to determine the charge / discharge power sequence within the predicted time period.

6. The method according to claim 5, characterized in that, The process of processing the target profit function using core decision variables and charging / discharging constraints to determine the charging / discharging power sequence within the prediction period includes: The market and carbon policy text data in the data to be used are processed based on a first language model to obtain carbon benefit weights; wherein, the carbon benefit weights include carbon benefit information corresponding to different time periods; Using the particle swarm optimization algorithm, the core decision variables and the charging / discharging constraints are processed to obtain multiple candidate power sequences; wherein, the target benefit function includes at least carbon benefit information; Based on the iterative update of the multiple candidate power sequences, the maximum value corresponding to the target profit function is determined, and the candidate power sequence corresponding to the maximum value is used as the charge and discharge power sequence within the prediction time.

7. The method according to claim 6, characterized in that, The method further includes: Generate an interpretable report for a preset object based on the charge and discharge power sequence within the predicted time period; The interpretable report shall include at least an overall revenue overview, time-segmented scheduling suggestions, risk warnings, and generation link information for generating the interpretable report, wherein the generation link information shall include at least the normative documents on which it is based.

8. A carbon emission reduction prediction device for energy storage based on a large language model, characterized in that, include: The data to be used module is used to acquire the data to be used generated in the energy storage battery during the emission period. The data to be used includes at least one or more of the following: numerical data of operation and measurement, structured data related to unit and emission, textual data of market and carbon policy, and data of external environmental load. The carbon emission and confidence attribute prediction module is used to determine the target feature tensor to be input into the target language model based on the data to be used, and input the target feature tensor into the pre-trained target language model to output the carbon emission and confidence attribute of the data to be used within the prediction period; wherein, the target feature tensor includes at least one or more of the charging power, discharging power, state of charge, target marginal emission factor, electricity trading attributes, environmental characteristics, and electricity policy documents at discrete time points.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the energy storage carbon emission reduction prediction method based on a large language model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the energy storage carbon emission reduction prediction method based on a large language model as described in any one of claims 1-7.

Images