Rebalancing Dynamic Adaptation Method and Service System Applied to Multi-Source Systems

Abstract

The present invention provides a rebalancing dynamic adjustment method and service system applied to a multi-source system, which relates to the field of smart energy technology. First, collect the real-time operation data set of the multi-source energy system, covering energy supply, load demand, and system operation state characteristics. Then, extract the dynamic supply-demand deviation characteristics from it to generate a characteristic set including supply, demand fluctuations, and system stability deviations. Next, call the pre-trained multi-source collaborative prediction model to generate a multi-source collaborative prediction result containing energy supply and load demand trend prediction characteristics based on the above characteristic set. Then, match the dynamic adjustment strategy according to the multi-source collaborative prediction result and the real-time operation data to obtain a strategy set such as energy output adjustment, energy storage charge and discharge, and load priority allocation. Finally, execute the adjustment strategy and obtain feedback data, and optimize the parameters of the multi-source collaborative prediction model based on this to trigger a closed-loop adjustment operation, realizing the rebalancing dynamic adjustment of the multi-source system.

Description

Technical Field

[0001] The present invention relates to the technical field of smart energy, and more particularly, to a rebalancing dynamic adjustment method and service system applied to a multi-source system. Background Art

[0002] In the field of smart energy, with the wide application of multi-source energy systems, their efficient and stable operation faces many challenges. Traditional management methods for multi-source energy systems often struggle to adapt to complex and changing operating environments, and obvious limitations have emerged in practical applications.

[0003] Most existing management methods for multi-source energy systems use simple statistical analysis methods, making it difficult to effectively extract dynamic supply-demand deviation characteristics from massive data, accurately identify key information such as supply fluctuations, demand fluctuations, and system stability deviations, and thus unable to provide strong support for subsequent decision-making. In terms of prediction, most existing prediction models are for single energy sources or simple scenarios, lacking effective prediction means for the coordinated operation of multi-source energy systems, unable to comprehensively consider the mutual influence and synergy between multiple energies, resulting in inaccurate prediction results and difficulty in reflecting the true trends of energy supply and load demand. In addition, traditional methods often rely on experience and fixed rules, lacking the ability to dynamically match and adjust according to real-time prediction results and system operating states, unable to generate a set of coordinated adjustment strategies in multiple aspects, and difficult to achieve comprehensive optimization of all links of the system. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a rebalancing dynamic adjustment method applied to a multi-source system, the method comprising:

[0005] Collect a real-time operation data set of the multi-source energy system, the real-time operation data set including energy supply characteristics, load demand characteristics, and system operation state characteristics;

[0006] Perform dynamic supply-demand deviation feature extraction processing on the real-time operation data set to generate a dynamic supply-demand deviation feature set including supply fluctuation characteristics, demand fluctuation characteristics, and system stability deviations;

[0007] Invoke a pre-trained multi-source collaborative prediction model to generate a multi-source collaborative prediction result based on the dynamic supply-demand deviation feature set, the multi-source collaborative prediction result including energy supply trend prediction characteristics and load demand trend prediction characteristics;

[0008] Perform dynamic adjustment strategy matching processing according to the multi-source collaborative prediction result and the real-time operation data set to generate a multi-source collaborative adjustment strategy set, the multi-source collaborative adjustment strategy set including energy output adjustment strategies, energy storage charge and discharge strategies, and load priority allocation strategies;

[0009] Execute the adaptation strategies in the multi-source collaborative adaptation strategy set and obtain adaptation execution feedback data, and optimize the parameters of the multi-source collaborative prediction model based on the adaptation execution feedback data to trigger a closed-loop adaptation operation.

[0010] In another aspect, an embodiment of the present invention further provides a multi-source energy debugging service system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiments of the present application achieve efficient, accurate and adaptive dynamic adaptation of the multi-source energy system, significantly improving the overall performance and operation efficiency of the multi-source energy system. Specifically, by collecting the real-time operation data set of the multi-source energy system, on this basis, performing dynamic supply-demand deviation feature extraction processing, identifying supply fluctuations, demand fluctuations and system stability deviations, calling the pre-trained multi-source collaborative prediction model, generating a multi-source collaborative prediction result based on the dynamic supply-demand deviation feature set, and performing dynamic adaptation strategy matching processing according to the multi-source collaborative prediction result and the real-time operation data set to generate a multi-source collaborative adaptation strategy set including energy output adjustment, energy storage charge and discharge, and load priority allocation, realizing comprehensive and collaborative optimization of all links of the system. Finally, execute the multi-source collaborative adaptation strategy and obtain adaptation execution feedback data, and optimize the parameters of the multi-source collaborative prediction model based on the adaptation execution feedback data to trigger a closed-loop adaptation operation, forming a complete feedback optimization mechanism, enabling the system to continuously learn and improve itself and adapt to the changing operation environment. Thereby, effectively solving the complex and changeable operation problems of the multi-source energy system, improving the energy utilization efficiency, and enhancing the stability and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic execution flow diagram of the rebalancing dynamic adaptation method applied to a multi-source system provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of exemplary hardware and software components of the multi-source energy debugging service system provided by an embodiment of the present invention;

[0014] Reference numerals in the drawings: 100 - multi-source energy debugging service system; 110 - network port; 120 - processor; 130 - bus; 140 - storage medium; 150 - I / O interface. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1It is a schematic flowchart of a rebalancing dynamic adjustment method applied to a multi-source system according to an embodiment of the present invention. The rebalancing dynamic adjustment method applied to the multi-source system will be introduced in detail below.

[0016] Step S110, collect the real-time operation data set of the multi-source energy system, where the real-time operation data set includes energy supply characteristics, load demand characteristics, and system operation state characteristics.

[0017] In this embodiment, there is a multi-source energy system in the energy park. This multi-source energy system integrates various energy supply methods and various load demands, aiming to achieve efficient energy distribution and stable supply. Specifically, in this energy park, the real-time operation data set covers energy supply characteristics, load demand characteristics, and system operation state characteristics. For example, in terms of energy supply, the photovoltaic power generation system starts working every day from sunrise in the morning, and its power usually changes with the change of light intensity. For example, at 9 am, the photovoltaic power generation power data is 500 kW, and it rises to 600 kW at 10 am. The above data forms a photovoltaic power generation power data sequence. The wind power generation system outputs power according to the wind speed. During a certain period of time, when the wind speed is relatively stable, the wind power generation power remains at about 800 kW. As the wind speed changes, the power will also fluctuate accordingly, forming a wind power generation power data sequence. The energy storage system performs charge and discharge operations according to the energy supply and demand situation, and its output power data records the power value and time of each charge and discharge. For example, at a certain moment, the energy storage system outputs electrical energy to the system at a power of 200 kW, thus forming an energy storage system output power data sequence.

[0018] In terms of load demand characteristics, there are many factories in the energy park, and the industrial load demand changes with the production plan and equipment operation conditions. When a certain factory conducts large-scale production in the morning, the industrial load demand data reaches 1500 kW at 10 am, and then gradually decreases as some equipment shuts down for maintenance, forming an industrial load demand data sequence. In the commercial area, the business hours of the stores and the usage conditions of the electrical equipment determine the commercial load demand. At about 12 noon, the commercial load demand data reaches 800 kW, forming a commercial load demand data sequence. The electricity consumption situation in the residential living area shows obvious periodicity. From 7 pm to 10 pm is the peak electricity consumption period, and the residential load demand data may rise from 600 kW to 900 kW during this period, forming a residential load demand data sequence.

[0019] In terms of the characteristics of the system operating state, the power grid frequency deviation data sequence reflects the difference between the actual operating frequency of the power grid and the standard frequency. For example, within a certain period of time, the power grid frequency deviation data fluctuates between -0.1 Hz and 0.2 Hz, and the above data is recorded to form the power grid frequency deviation data sequence. The voltage deviation data sequence reflects the deviation of the power grid voltage from the rated voltage. For example, at a certain moment, the voltage deviation data is +5 V, and the voltage deviation data at different moments is recorded to form the voltage deviation data sequence. That is, through data acquisition devices and sensors, the above various types of data are collected in real time to form a real-time operation data set of the multi-source energy system.

[0020] Step S120: Perform dynamic supply-demand deviation feature extraction processing on the real-time operation data set to generate a dynamic supply-demand deviation feature set including supply fluctuation features, demand fluctuation features, and system stability deviations.

[0021] In this embodiment, time series fluctuation analysis can be performed on the photovoltaic power generation power data sequence in the energy supply characteristics. For example, the instantaneous power change rate can be calculated at a time interval of 10 minutes. For example, from 9:00 am to 9:10 am, the photovoltaic power generation power rises from 500 kW to 550 kW, and its instantaneous power change rate is (550 - 500) ÷ 500 × 100% = 10%, and this 10% can be understood as a part of the first supply fluctuation sub-feature. For the wind power generation power data sequence, calculate its power standard deviation. Suppose within a certain hour, the wind power generation power data is 780 kW, 820 kW, 800 kW, etc., and the calculated power standard deviation is 20 kW, and this 20 kW can be understood as the second supply fluctuation sub-feature. For the energy storage system output power data sequence, count its charge-discharge switching frequency. Within a day, the energy storage system switches 15 times between charge and discharge, so the charge-discharge switching frequency is 15 times / day, which can be used as the third supply fluctuation sub-feature.

[0022] Moreover, demand mutation detection processing can be performed on the industrial load demand data sequence in the load demand characteristics. For example, through data analysis, it is detected that at 11:00 am, due to the startup of a certain large equipment, the industrial load demand suddenly rises from 1200 kW to 1800 kW, and this 11:00 am is the mutation time point, and all such mutation time points are recorded to form the first demand fluctuation sub-feature. For the commercial load demand data sequence, analyze its load fluctuation amplitude. For example, within a day, the commercial load demand ranges from a minimum of 500 kW to a maximum of 900 kW, and the load fluctuation amplitude is 900 - 500 = 400 kW, which is used as the second demand fluctuation sub-feature. For the residential load demand data sequence, through the analysis of multi-day data, it is found that the residential electricity consumption shows a periodic fluctuation pattern with a peak from 7:00 pm to 10:00 pm and a trough from 2:00 am to 5:00 am every day, which constitutes the third demand fluctuation sub-feature.

[0023] Further, the stability index of the grid frequency deviation data sequence in the system operation state characteristics can be calculated. The grid frequency deviation data within a certain period of time is accumulated. For example, within 8 hours, the grid frequency deviation data are -0.1Hz, 0Hz, 0.1Hz, etc., and the cumulative offset is -0.1 + 0 + 0.1 = 0Hz, which is used as the first stability deviation sub - feature. For the voltage deviation data sequence, its instantaneous fluctuation extreme value is found. At a certain moment, the voltage deviation reaches +8V, which is the second stability deviation sub - feature. Finally, all the above sub - features are combined to generate a dynamic supply - demand deviation feature set including supply fluctuation features, demand fluctuation features, and system stability deviation.

[0024] Step S130, call the pre - trained multi - source collaborative prediction model, and generate a multi - source collaborative prediction result based on the dynamic supply - demand deviation feature set. The multi - source collaborative prediction result includes an energy supply trend prediction feature and a load demand trend prediction feature.

[0025] In this embodiment, the dynamic supply - demand deviation feature set can be input into the encoder layer of the multi - source collaborative prediction model. This encoder layer can perform multi - dimensional feature fusion processing on the dynamic supply - demand deviation feature set. For example, the supply fluctuation features such as the instantaneous power change rate of photovoltaic power generation, the standard deviation of wind power generation, and the charge - discharge switching frequency of the energy storage system are fused with the demand fluctuation features such as the industrial load demand mutation time point, the commercial load demand fluctuation amplitude, and the residential load demand periodic fluctuation pattern, as well as the system stability deviation features such as the cumulative offset of grid frequency deviation and the instantaneous fluctuation extreme value of voltage deviation. After fusion, a fusion feature vector including supply - demand coupling features and system stability correlation features is generated.

[0026] Next, the time series prediction layer of the multi-source collaborative prediction model is called to perform time series extension processing on the fused feature vector. For example, the energy supply trend and load demand trend within the next 24 hours can be predicted. Among them, the energy supply trend prediction features include the predicted curve of photovoltaic power generation. It is predicted that within the next 24 hours, the photovoltaic power generation will reach a peak of 1,200 kW around 12:00 noon and then gradually decrease. The predicted curve of wind power generation shows that from 8:00 pm to 10:00 pm, as the wind speed increases, the wind power generation will rise from 700 kW to 900 kW. The predicted curve of the dispatchable capacity of the energy storage system indicates that at 4:00 pm, the dispatchable capacity of the energy storage system will reach 500 kWh. The load demand trend prediction features include the predicted curve of industrial load demand. It is predicted that from 2:00 pm to 4:00 pm, the industrial load demand will rise from 1,600 kW to 1,800 kW. The predicted curve of commercial load demand shows that from 7:00 pm to 9:00 pm, the commercial load demand will rise from 700 kW to 900 kW. The predicted curve of residential load demand indicates that from 8:00 pm to 10:00 pm, the residential load demand will rise from 800 kW to 1,000 kW.

[0027] Then, the calculation of the supply-demand matching degree can be performed between the energy supply trend prediction features and the load demand trend prediction features. For example, it is calculated that at 9:00 pm, there is a supply-demand gap of 200 kW between the energy supply and the load demand, which is the predicted value of the supply-demand gap. At the same time, it is predicted that around 11:00 pm, the energy supply and the load demand will reach balance, which is the supply-demand balance time point. By combining the above results, a multi-source collaborative prediction result including the energy supply trend prediction features and the load demand trend prediction features can be generated.

[0028] Step S140, perform dynamic adaptation strategy matching processing according to the multi-source collaborative prediction result and the real-time operation data set to generate a multi-source collaborative adaptation strategy set, where the multi-source collaborative adaptation strategy set includes an energy output adjustment strategy, an energy storage charge and discharge strategy, and a load priority allocation strategy.

[0029] In this embodiment, the predicted value of the supply-demand gap can be matched with a plurality of preset supply-demand balance threshold intervals. Suppose the preset threshold intervals are: [-300 kW, -100 kW], [-100 kW, 100 kW], [100 kW, 300 kW], etc. The predicted value of the supply-demand gap at 9:00 pm is 200 kW, which belongs to the target threshold interval of [100 kW, 300 kW].

[0030] Then, according to the strategy generation rules corresponding to the target threshold interval, the set of dispatchable energy supply sources can be screened from the energy supply characteristics. For example, since the photovoltaic power generation system has basically stopped working at 9 o'clock in the evening, the wind power generation system still has a certain adjustment space, and the energy storage system can also perform charging and discharging operations, the set of dispatchable energy supply sources includes wind power generation systems and energy storage systems. Based on the current state of charge and dispatchable capacity in the energy storage charging and discharging strategy, assuming that the current state of charge of the energy storage system is 60% and the dispatchable capacity is 400 kWh, it is calculated that the maximum charging and discharging power of the energy storage system is 150 kW.

[0031] Next, according to the preset industry priority weights in the load priority allocation strategy, industrial loads have a higher priority due to production continuity requirements; commercial loads are second; and residential loads have a relatively lower priority. The industrial load demand data series, commercial load demand data series, and residential load demand data series are prioritized for load reduction. For example, when load reduction is required, residential loads are prioritized, followed by commercial loads, and the stable operation of industrial loads is guaranteed as much as possible.

[0032] Finally, based on the set of dispatchable energy supply sources, maximum charging and discharging power, and load reduction priority ranking, a set of multi-source coordinated adaptation strategies is generated. The energy output adjustment strategy is to appropriately increase the output power of the wind power generation system, for example, to increase the wind power generation power from the current 750 kilowatts to 850 kilowatts; the energy storage charging and discharging strategy is to allow the energy storage system to discharge at a power of 100 kilowatts; the load priority allocation strategy is to appropriately reduce the residential load when necessary, such as reducing the residential load by 100 kilowatts.

[0033] Step S150, executing the adaptation strategy in the multi-source collaborative adaptation strategy set and obtaining adaptation execution feedback data, and optimizing the parameters of the multi-source collaborative prediction model based on the adaptation execution feedback data to trigger a closed-loop adaptation operation.

[0034] In this embodiment, a power adjustment instruction in the energy output adjustment strategy can be sent to the photovoltaic inverter controller of the multi-source energy system. Since the power of the photovoltaic power generation system is relatively low at this time, the wind power generation system is mainly adjusted. After receiving the power adjustment instruction, the wind power generation system increases the power from 750 kilowatts to 850 kilowatts, and the actual adjustment response data of the photovoltaic power generation power is obtained (although the photovoltaic power generation power is not adjusted, the relevant data is recorded for overall feedback).

[0035] In addition, the charge and discharge rate instructions in the energy storage charge and discharge strategy can be sent to the bidirectional converter of the energy storage system, and the energy storage system discharges at a power of 100 kilowatts to obtain the actual charge and discharge power data of the energy storage system. For example, within 1 hour after executing the instruction, the energy storage system actually discharges at an average power of 95 kilowatts.

[0036] Moreover, a load shedding instruction in the load priority allocation strategy can also be sent to the load management terminal to shed the residential load and obtain the actual load shedding amount data. For example, 80 kW of residential load is actually shed.

[0037] Furthermore, the actual adjustment response data, the actual charge and discharge power data, and the actual load shedding amount data can be combined to generate the adjustment execution feedback data. Calculate the policy execution deviation index between the adjustment execution feedback data and the multi-source collaborative prediction result. For example, in the multi-source collaborative prediction result, it is predicted that the energy storage system discharges at a power of 100 kW, and the actual average power is 95 kW, with a deviation of 5 kW; it is predicted to shed 100 kW of residential load, and the actual load shedding is 80 kW, with a deviation of 20 kW, etc. The policy execution deviation index is obtained through comprehensive calculation.

[0038] Then, based on the adjustment execution feedback data, the parameters of the multi-source collaborative prediction model are optimized to trigger the closed-loop adjustment operation. Then, the policy execution deviation index is input into the feedback learning layer of the multi-source collaborative prediction model, the gradient change amount of the feature fusion weight matrix in the encoder layer is calculated according to the policy execution deviation index, and the feature fusion weight matrix is updated based on the gradient descent algorithm. For example, when it is detected that the weight settings for fusing certain supply fluctuation features and demand fluctuation features are unreasonable, resulting in a deviation between the prediction result and the actual execution situation, the weight matrix is adjusted so that the above features more accurately reflect the actual situation when fusing.

[0039] Finally, according to the time dimension deviation component in the policy execution deviation index, the window length and the sliding step length in the time window expansion parameter of the time series prediction layer are adjusted. If the time dimension deviation component increases, it indicates that the time accuracy of the prediction decreases. The window length is shortened and the sliding step length is increased. For example, the window length is shortened from the original 6 hours to 4 hours, and the sliding step length is increased from 1 hour to 2 hours; if the time dimension deviation component decreases, the window length is extended and the sliding step length is decreased.

[0040] Thus, the updated multi-source collaborative prediction result can be regenerated according to the adjusted feature fusion weight matrix and the time window expansion parameter, and the updated multi-source collaborative prediction result is subjected to a secondary dynamic adjustment strategy matching process with the real-time operation data set to generate an optimized multi-source collaborative adjustment strategy set. By repeatedly executing the optimized multi-source collaborative adjustment strategy set and iteratively updating the parameters of the multi-source collaborative prediction model until the policy execution deviation index is lower than the preset deviation threshold, the efficient and stable operation of the multi-source energy system is achieved.

[0041] Based on the above steps, the embodiments of the present application achieve efficient, precise, and adaptive dynamic adjustment of the multi-source energy system, significantly improving the overall performance and operating efficiency of the multi-source energy system. Specifically, by collecting the real-time operation data set of the multi-source energy system, on this basis, dynamic supply-demand deviation feature extraction processing is performed to identify supply fluctuations, demand fluctuations, and system stability deviations. The pre-trained multi-source collaborative prediction model is called to generate multi-source collaborative prediction results based on the dynamic supply-demand deviation feature set. According to the multi-source collaborative prediction results and the real-time operation data set, dynamic adjustment strategy matching processing is carried out to generate a multi-source collaborative adjustment strategy set including energy output adjustment, energy storage charging and discharging, and load priority allocation, realizing the comprehensive and collaborative optimization of each link of the system. Finally, the multi-source collaborative adjustment strategy is executed and the adjustment execution feedback data is obtained, and the parameters of the multi-source collaborative prediction model are optimized based on the adjustment execution feedback data to trigger the closed-loop adjustment operation, forming a complete feedback optimization mechanism, enabling the system to continuously self-learn and improve, and adapt to the changing operating environment. Thus, the complex and variable operation problems of the multi-source energy system are effectively solved, the energy utilization efficiency is improved, and the stability and reliability of the system are enhanced.

[0042] In a possible implementation manner, step S120 includes:

[0043] Step S121, perform time series fluctuation analysis on the photovoltaic power data sequence, wind power data sequence, and energy storage system output power data sequence in the energy supply characteristics, extract the instantaneous power change rate of the photovoltaic power data sequence as the first supply fluctuation sub-feature, extract the power standard deviation of the wind power data sequence as the second supply fluctuation sub-feature, and extract the charge-discharge switching frequency of the energy storage system output power data sequence as the third supply fluctuation sub-feature.

[0044] In this embodiment, for the photovoltaic power data sequence, starting from 8:00 in the morning, the power value is recorded every 15 minutes. The power at 8:00 is 300 kW, and the power at 8:15 becomes 350 kW. Calculate the instantaneous power change rate. Subtract the power at 8:00 from the power at 8:15, that is, 350 kW minus 300 kW, to get a power change amount of 50 kW. Then divide the power change amount by the power at 8:00, that is, 50 kW divided by 300 kW, which is approximately equal to 0.167, and converted to a percentage is 16.7%. This 16.7% is the instantaneous power change rate of the photovoltaic power data sequence in this period, as the first supply fluctuation sub-feature.

[0045] For the wind power generation power data sequence, within a certain day, the wind power generation power is recorded once per hour, and the data are 750 kW, 800 kW, 780 kW, 820 kW, and 790 kW respectively. First, calculate the average value of this set of data. Add the above data: 750 + 800 + 780 + 820 + 790 = 3940 kW, and then divide by the number of data, which is 5. The average value is 3940 kW divided by 5, equal to 788 kW. Then calculate the square of the difference between each data and the average value: (750 - 788)² = (-38)² = 1444, (800 - 788)² = 12² = 144, (780 - 788)² = (-8)² = 64, (820 - 788)² = 32² = 1024, (790 - 788)² = 2² = 4. Next, find the average value of the above square values: (1444 + 144 + 64 + 1024 + 4) ÷ 5 = 2680 ÷ 5 = 536. Finally, take the square root of this average value, and the power standard deviation is approximately 23.15 kW, which is the second supply fluctuation sub-feature.

[0046] For the energy storage system output power data sequence, within one day, the number of charge-discharge switching times of the energy storage system can be counted. For example, from 7 am to 7 pm, a total of 12 charge-discharge switchings of the energy storage system are recorded. Then the charge-discharge switching frequency is 12 times / 12 hours, that is, 1 time / hour, which is the third supply fluctuation sub-feature.

[0047] Step S122: Perform demand mutation detection processing on the industrial load demand data sequence, commercial load demand data sequence, and residential load demand data sequence in the load demand characteristics, determine the set of mutation time points of the industrial load demand data sequence as the first demand fluctuation sub-feature, determine the load fluctuation amplitude of the commercial load demand data sequence as the second demand fluctuation sub-feature, and determine the periodic fluctuation pattern of the residential load demand data sequence as the third demand fluctuation sub-feature.

[0048] In this embodiment, for the industrial load demand data sequence, for example, at 10 am in a certain factory, due to the start of a new production line, the load demand suddenly rises from 1000 kW to 1500 kW, and the time point of 10 am is recorded. Then at 2 pm, due to partial equipment failures and shutdowns, the load demand drops from 1400 kW to 1000 kW, and 2 pm is also recorded. Thus, the above set of mutation time points can be gathered as the first demand fluctuation sub-feature.

[0049] For the commercial load demand data series, within a certain week, the maximum and minimum values of the commercial load demand are recorded every day. For example, the minimum value on Monday is 600 kW and the maximum value is 900 kW; the minimum value on Tuesday is 580 kW and the maximum value is 880 kW, etc. By comparing the above data, the maximum fluctuation range of the commercial load demand within a week is determined, that is, the maximum value minus the minimum value. For example, 900 kW minus 580 kW equals 320 kW, and this 320 kW is the second demand fluctuation sub - feature.

[0050] For the residential load demand data series, the residential load demand data for each day within a month can be analyzed. For example, it is found that from 7 pm to 10 pm every day is the peak electricity consumption period, and the load demand gradually rises from 500 kW at 7 pm to 800 kW at 10 pm, and then gradually decreases. From 2 am to 4 am is the low - electricity - consumption period, and the load demand remains at about 300 kW. This daily repetitive periodic fluctuation pattern is used as the third demand fluctuation sub - feature.

[0051] Step S123, calculate the stability index for the power grid frequency deviation data series and the voltage deviation data series in the system operation state characteristics, obtain the cumulative offset of the power grid frequency deviation as the first stability deviation sub - feature, and obtain the instantaneous fluctuation extreme value of the voltage deviation as the second stability deviation sub - feature.

[0052] In this embodiment, for the power grid frequency deviation data series, the power grid frequency deviation value can be recorded every half an hour. For example, within one day, the recorded data are - 0.05 Hz, 0 Hz, 0.03 Hz, - 0.02 Hz, etc. By cumulatively adding the above data, - 0.05 + 0 + 0.03+( - 0.02)= - 0.04 Hz, and this - 0.04 Hz is the cumulative offset of the power grid frequency deviation, serving as the first stability deviation sub - feature.

[0053] For the voltage deviation data series, within a certain day, the voltage deviation value is monitored in real - time. The maximum value recorded is + 6 V and the minimum value is - 4 V, where + 6 V is the instantaneous fluctuation extreme value of the voltage deviation, serving as the second stability deviation sub - feature.

[0054] Step S124, merge the first supply fluctuation sub - feature, the second supply fluctuation sub - feature, the third supply fluctuation sub - feature, the first demand fluctuation sub - feature, the second demand fluctuation sub - feature, the third demand fluctuation sub - feature, the first stability deviation sub - feature and the second stability deviation sub - feature to generate a dynamic supply - demand deviation feature set.

[0055] In this embodiment, the first supply fluctuation sub-feature (the instantaneous power change rate of the photovoltaic power generation power data sequence is 16.7%), the second supply fluctuation sub-feature (the power standard deviation of the wind power generation power data sequence is 23.15 kW), the third supply fluctuation sub-feature (the charge and discharge switching frequency of the energy storage system output power data sequence is 1 time / hour), the first demand fluctuation sub-feature (the set of mutation time points of the industrial load demand data sequence, such as 10 am, 2 pm, etc.), the second demand fluctuation sub-feature (the load fluctuation amplitude of the commercial load demand data sequence is 320 kW), the third demand fluctuation sub-feature (the periodic fluctuation pattern of the residential load demand data sequence, with peaks from 7 pm to 10 pm every night and valleys from 2 am to 4 am), the first stability deviation sub-feature (the cumulative offset of the power grid frequency deviation is -0.04 Hz), and the second stability deviation sub-feature (the instantaneous fluctuation extreme value of the voltage deviation is +6V) can be combined to generate a dynamic supply-demand deviation feature set.

[0056] In a possible implementation manner, step S130 includes:

[0057] Step S131, input the dynamic supply-demand deviation feature set into the encoder layer of the multi-source collaborative prediction model, and perform multi-dimensional feature fusion processing on the dynamic supply-demand deviation feature set through the encoder layer to generate a fusion feature vector including supply-demand coupling features and system stability correlation features.

[0058] For example, the instantaneous power change rate of the photovoltaic power generation power data sequence in the energy supply fluctuation feature is 15%, the power standard deviation of the wind power generation power data sequence is 25 kW, and the charge and discharge switching frequency of the energy storage system output power data sequence is 1.5 times / hour; in the load demand fluctuation feature, the set of mutation time points of the industrial load demand data sequence is 11 am and 3 pm, the load fluctuation amplitude of the commercial load demand data sequence is 350 kW, and the periodic fluctuation pattern of the residential load demand data sequence is that there are peaks from 7 pm to 10 pm every night and valleys from 2 am to 4 am; in the system stability deviation feature, the cumulative offset of the power grid frequency deviation is -0.06 Hz, and the instantaneous fluctuation extreme value of the voltage deviation is +8V.

[0059] The encoder layer performs multi-dimensional feature fusion processing on the above dynamic supply-demand deviation features. For example, the mutual relationship between energy supply and load demand can be comprehensively considered, as well as the impact of system stability on them. For example, the instantaneous power change rate of photovoltaic power generation is combined with the mutation time point of industrial load demand to analyze whether there is a potential correlation between the two. It may be found that when the photovoltaic power generation power rises rapidly, the startup of some industrial equipment will cause a mutation in industrial load demand. Through this correlation analysis, partial information containing supply-demand coupling characteristics is generated. At the same time, the cumulative offset of the grid frequency deviation is correlated with the standard deviation of the wind power generation power, considering the impact of system stability on the fluctuation of energy supply, so as to generate system stability correlation characteristics. Thus, the above information after fusion processing together constitutes a fusion feature vector.

[0060] Step S132: Invoke the time series prediction layer of the multi-source collaborative prediction model to perform time series extension processing on the fusion feature vector, and generate energy supply trend prediction features and load demand trend prediction features within a future preset time window. Among them, the energy supply trend prediction features include a photovoltaic power generation power prediction curve, a wind power generation power prediction curve, and a schedulable capacity prediction curve of the energy storage system, and the load demand trend prediction features include an industrial load demand prediction curve, a commercial load demand prediction curve, and a residential load demand prediction curve.

[0061] In this embodiment, the future 24 hours is used as the preset time window to generate energy supply trend prediction features and load demand trend prediction features.

[0062] In terms of the energy supply trend prediction features, for the photovoltaic power generation power prediction curve, by analyzing historical data and current factors such as weather conditions and time, it is predicted that within the next 24 hours, the photovoltaic power generation power will change with the change of sunlight intensity. For example, there will be a certain power output starting at around 7 am. As the sunlight intensifies, the power gradually rises, and it is expected to reach a peak of 1300 kW at around 12 noon. Then, as the sunlight weakens, the power gradually decreases and basically stops generating electricity at around 7 pm.

[0063] The wind power generation power prediction curve takes into account current factors such as wind speed, wind direction, and the operating state of wind power generation equipment. It is predicted that within the next 24 hours, due to changes in meteorological conditions, the wind speed will increase from 9 pm to 11 pm, and the wind power generation power will gradually rise from 800 kW at 9 pm to 950 kW at 11 pm. Then, as the wind speed decreases, the power also gradually decreases.

[0064] The dispatchable capacity prediction curve of the energy storage system is generated based on the current state of charge of the energy storage system, the historical charge and discharge data, and the predicted future energy supply and demand. Suppose the current state of charge of the energy storage system is 55% and the dispatchable capacity is 450 kWh. Considering the changes in future energy supply and load demand, it is predicted that within the next 24 hours, the energy storage system will charge when the photovoltaic power generation is excessive and discharge when the load demand is at a peak and the energy supply is insufficient. It is expected that from 4 pm to 6 pm, the energy storage system will charge at a power of 120 kW, and the dispatchable capacity will reach 580 kWh at 6 pm; from 8 pm to 10 pm, the energy storage system will discharge at a power of 150 kW, and the dispatchable capacity will drop to 330 kWh at 10 pm.

[0065] In terms of the prediction characteristics of the load demand trend, the industrial load demand prediction curve is generated based on the production plans and equipment operation arrangements of each factory in the park. For example, a large factory plans to carry out large-scale production activities from 2 pm to 6 pm. It is predicted that the industrial load demand will gradually increase from 1600 kW at 2 pm to 1900 kW at 4 pm, and then decrease to 1700 kW at 6 pm as some production links are completed.

[0066] The commercial load demand prediction curve combines the business hours of stores and the usage patterns of electrical equipment in the commercial area. It is expected that from 7 pm to 9 pm, as the passenger flow in commercial places such as shopping malls and restaurants increases, the commercial load demand will gradually increase from 750 kW at 7 pm to 900 kW at 9 pm, and then gradually decrease as the stores close one after another.

[0067] The residential load demand prediction curve is based on the daily electricity consumption habits of residents. From 7 pm to 10 pm is the peak period of residential electricity consumption. It is predicted that the residential load demand will gradually increase from 700 kW at 7 pm to 950 kW at 9 pm, and then decrease to 600 kW at 11 pm as residents go to sleep.

[0068] Step S133: Calculate the supply-demand matching degree between the energy supply trend prediction characteristics and the load demand trend prediction characteristics to generate a multi-source collaborative prediction result including the predicted value of the supply-demand gap and the supply-demand balance time point.

[0069] For example, at 9 pm, the photovoltaic power generation is basically 0 kW, the wind power generation is 900 kW, the power output corresponding to the dispatchable capacity of the energy storage system is 120 kW, and the total energy supply is 900 + 120 = 1020 kW; while at this time, the industrial load demand is 1750 kW, the commercial load demand is 850 kW, and the residential load demand is 900 kW, and the total load demand is 1750 + 850 + 900 = 3500 kW. Subtracting the two, the predicted value of the supply-demand gap is 3500 - 1020 = 2480 kW.

[0070] On this basis, continue to analyze the energy supply and load demand at each time point to find the time point of supply-demand balance. Through calculation and analysis, it is found that around 3 am, due to a significant reduction in industrial load demand, the commercial load demand is basically 0, and the residential load demand is also at a relatively low level. At this time, the wind power generation still has a certain power output, and the energy storage system also has a certain schedulable capacity, and the energy supply and load demand reach balance. This 3 am is the time point of supply-demand balance. Combining information such as the predicted value of the supply-demand gap and the time point of supply-demand balance generates a multi-source collaborative prediction result that includes the predicted characteristics of the energy supply trend and the predicted characteristics of the load demand trend.

[0071] In one possible implementation, step S140 includes:

[0072] Step S141, match the predicted value of the supply-demand gap with a plurality of preset supply-demand balance threshold intervals to determine the target threshold interval to which the current supply-demand gap belongs.

[0073] Suppose the preset threshold intervals are: [-3000 kW, -2000 kW], [-2000 kW, -1000 kW], [-1000 kW, 1000 kW], [1000 kW, 2000 kW], [2000 kW, 3000 kW], etc. The predicted value of the supply-demand gap at 9 pm is 2480 kW, which belongs to the target threshold interval of [2000 kW, 3000 kW].

[0074] Step S142, according to the policy generation rule corresponding to the target threshold interval, screen the set of schedulable energy supply sources from the energy supply characteristics, and determine the maximum charge-discharge power of the energy storage system based on the current state of charge and schedulable capacity in the energy storage charge-discharge strategy.

[0075] For example, at 9 pm, the photovoltaic power generation system basically stops working and is not schedulable; although the wind power generation system has a certain power output, the improvement space is limited; while the energy storage system still has a certain schedulable capacity. So the set of schedulable energy supply sources is mainly the energy storage system.

[0076] Suppose the current state of charge of the energy storage system is 40% and the schedulable capacity is 300 kWh. Considering factors such as the safe operation and charge-discharge efficiency of the energy storage system, after a series of calculations and evaluations (such as combining the rated power of the energy storage system, charge-discharge historical data, and current environmental factors such as temperature), the maximum charge-discharge power of the energy storage system is determined to be 180 kW.

[0077] Step S143: According to the preset industry priority weights in the load priority allocation strategy, perform load shedding priority sorting on the industrial load demand data sequence, the commercial load demand data sequence, and the residential load demand data sequence.

[0078] For example, due to the requirement of production continuity, the preset industrial load has the highest priority; the commercial load comes second; the residential load has a relatively lower priority. Therefore, when load shedding is needed, the residential load is considered for shedding first, followed by the commercial load, and every effort is made to ensure the stable operation of the industrial load.

[0079] Step S144: Based on the set of schedulable energy supply sources, the maximum charge-discharge power, and the load shedding priority sorting, generate a multi-source collaborative adjustment strategy set including an energy output adjustment strategy, a energy storage charge-discharge strategy, and a load priority allocation strategy.

[0080] For example, based on the set of schedulable energy supply sources (energy storage system), the maximum charge-discharge power (180 kW), and the load shedding priority sorting, generate a multi-source collaborative adjustment strategy set including an energy output adjustment strategy, a energy storage charge-discharge strategy, and a load priority allocation strategy. In terms of the energy output adjustment strategy, since the schedulable energy at this time is mainly the energy storage system, the strategy is to let the energy storage system discharge at the maximum charge-discharge power of 180 kW to increase energy supply. The energy storage charge-discharge strategy is clearly defined as continuously discharging at a power of 180 kW for a period of time until the state of charge of the energy storage system reaches the safety lower limit or the supply-demand gap is effectively alleviated. The load priority allocation strategy is to give priority to appropriately reducing the residential load. For example, during the peak electricity consumption period from 9 pm to 10 pm, reduce the residential load by 200 kW; if there is still a supply-demand gap, then consider reducing part of the commercial load, such as reducing the commercial load by 150 kW, to ensure that the industrial load can operate normally and the production activities in the park are not affected too much. Through the formulation and implementation of such a series of strategies, a complete multi-source collaborative adjustment strategy set is formed to cope with the real-time supply-demand imbalance in the multi-source energy system and ensure the stable supply and efficient utilization of energy.

[0081] In a possible implementation manner, step S150 includes:

[0082] Step S151: Send the power adjustment instruction in the energy output adjustment strategy to the photovoltaic inverter controller of the multi-source energy system, and obtain the actual adjustment response data of the photovoltaic power generation.

[0083] In this embodiment, since the photovoltaic power generation system has basically stopped working at this time, although a power adjustment instruction is sent, the photovoltaic power generation cannot be effectively adjusted in practice. The actual adjustment response data of the photovoltaic power generation finally obtained shows that the photovoltaic power generation remains at the current low level, such as 50 kW, and does not change due to the instruction.

[0084] Step S152: Send the charge and discharge rate instruction in the energy storage charge and discharge strategy to the bidirectional converter of the energy storage system, and obtain the actual charge and discharge power data of the energy storage system.

[0085] In this embodiment, after receiving the charge and discharge rate instruction, the energy storage system starts to operate at a set charge and discharge power of 180 kW. In the next hour, the actual charge and discharge power data of the energy storage system is obtained through real-time monitoring equipment. After recording and calculation, the energy storage system actually discharges at an average power of 175 kW in this hour. The specific calculation process is as follows: The discharge power of the energy storage system is recorded every 10 minutes, which are 170 kW, 172 kW, 178 kW, 176 kW, 174 kW, and 175 kW respectively. The sum of the above data is 1025 kW, and then divided by the number of records 6, the average discharge power is 1025 kW ÷ 6 ≈ 175 kW.

[0086] Step S153: Send the load shedding instruction in the load priority allocation strategy to the load management terminal, and obtain the actual load shedding amount data.

[0087] For example, according to the load priority allocation strategy, the residential load is shed first. After receiving the instruction, the load management terminal adjusts the residential load. After statistics, 180 kW of residential load is actually shed. During the shedding process, the electricity consumption changes of each residential user are recorded, and the actual load shedding amount is obtained by summarizing the above data.

[0088] Step S154: Combine the actual adjustment response data, the actual charge and discharge power data, and the actual load shedding amount data to generate the adjustment execution feedback data, and calculate the strategy execution deviation index between the adjustment execution feedback data and the multi-source collaborative prediction result.

[0089] For example, the actual adjusted response data (photovoltaic power generation maintained at 50 kW), actual charge and discharge power data (the energy storage system discharges at an average power of 175 kW), and actual load reduction data (180 kW of residential load is actually reduced) are combined to generate adaptation execution feedback data. Next, the strategy execution deviation index between the adaptation execution feedback data and the multi-source collaborative prediction result is calculated. In the multi-source collaborative prediction result, it is predicted that the energy storage system discharges at a power of 180 kW, and the actual average power is 175 kW, with a deviation of 180 - 175 = 5 kW; it is predicted to reduce 200 kW of residential load, and the actual reduction is 180 kW, with a deviation of 200 - 180 = 20 kW. Considering the deviations in these two aspects comprehensively, through a specific calculation method (for example, calculating after assigning weights according to the importance of different factors, assuming the weight of the energy storage system discharge power deviation is 0.6, and the weight of the residential load reduction deviation is 0.4), the strategy execution deviation index is calculated. The specific calculation is as follows: the influence value of the energy storage system discharge power deviation is 5 kW × 0.6 = 3, the influence value of the residential load reduction deviation is 20 kW × 0.4 = 8, and the sum of the two gives the strategy execution deviation index of 3 + 8 = 11.

[0090] Step S155, input the strategy execution deviation index into the feedback learning layer of the multi-source collaborative prediction model, and dynamically adjust the feature fusion weight matrix of the encoder layer and the time window expansion parameter of the time series prediction layer through the feedback learning layer.

[0091] For example, the strategy execution deviation index 11 can be input into the feedback learning layer of the multi-source collaborative prediction model. The feedback learning layer starts to dynamically adjust the feature fusion weight matrix of the encoder layer and the time window expansion parameter of the time series prediction layer.

[0092] For example, through analysis, it is found that in the previous feature fusion process, the weight settings of some features related to the energy storage system and load demand were not accurate enough, resulting in a deviation between the prediction result and the actual execution situation. For example, the weight of the feature of the charge and discharge switching frequency of the energy storage system was relatively low during fusion, and its impact on the overall supply and demand situation was not fully reflected. The feature fusion weight matrix is updated based on the gradient descent algorithm. The specific operation is to gradually adjust the weights of each feature according to the deviation index and a preset learning rate (for example, the learning rate is 0.01). Assume that the original weight of the feature of the charge and discharge switching frequency of the energy storage system is 0.2. After calculation, according to the gradient descent algorithm, the new weight is 0.2 + 0.01 × (the influence value of the deviation on the weight of this feature). Through complex calculations (involving the internal algorithms and data relationships of the model), the new weight is obtained as 0.23. Similar adjustments are made to the weights of other features, thus completing the update of the feature fusion weight matrix.

[0093] For the time window extension parameter of the time series prediction layer, it is adjusted according to the time dimension deviation component in the policy execution deviation index. Analysis shows that there is a certain deviation in the time dimension when predicting the charge and discharge power of the energy storage system and the load curtailment amount, indicating that the setting of the time window extension parameter is not very reasonable. The time dimension deviation component in the policy execution deviation index indicates that the time accuracy of the prediction needs to be improved. The current time window length is 6 hours, and the sliding step is 1 hour. Due to the increase in the time dimension deviation component, the window length is shortened and the sliding step is increased according to the rules. The window length is shortened to 4 hours, and the sliding step is increased to 2 hours. After such adjustment, the time series prediction layer can more accurately capture the changing trends of energy supply and load demand in the time dimension.

[0094] Step S156, regenerate the updated multi-source collaborative prediction result according to the adjusted feature fusion weight matrix and time window extension parameter, and perform a secondary dynamic adaptation strategy matching process on the updated multi-source collaborative prediction result and the real-time operation data set to generate an optimized multi-source collaborative adaptation strategy set.

[0095] In this embodiment, the updated multi-source collaborative prediction result is more accurate in both energy supply trend prediction and load demand trend prediction. For example, the prediction of the dispatchable capacity of the energy storage system is closer to the actual situation, and the prediction of industrial, commercial, and residential load demands in different time periods is also more accurate.

[0096] Suppose the newly calculated and analyzed predicted value of the supply-demand gap is within another target threshold range [-1000 kW, -500 kW]. According to the policy generation rules corresponding to the new target threshold range, re-screen the set of dispatchable energy supply sources from the energy supply characteristics. At this time, it is found that in addition to the energy storage system, the wind power generation system also has a certain dispatchable space. Based on the current state of charge and dispatchable capacity in the energy storage charge and discharge strategy, re-determine the maximum charge and discharge power of the energy storage system. After calculation (considering factors such as the current state of charge, charge and discharge efficiency, and remaining life of the energy storage system), it is determined that the maximum charge and discharge power of the energy storage system is 150 kW. At the same time, according to the preset industry priority weights in the load priority allocation strategy, re-rank the load curtailment priorities of the industrial load demand data sequence, commercial load demand data sequence, and residential load demand data sequence. Since the supply-demand gap is relatively small at this time, the load curtailment strategy may be adjusted, for example, appropriately reducing the residential load curtailment amount and increasing the regulation intensity of the commercial load.

[0097] Then, based on the re-screened set of schedulable energy supply sources (energy storage system and wind power generation system), the newly determined maximum charge-discharge power (150 kW for the energy storage system), and the re-ordered load shedding priorities, an optimized multi-source collaborative adjustment strategy set is generated. The energy output adjustment strategy is to appropriately increase the output power of the wind power generation system while controlling the energy storage system to perform charge-discharge operations at a power of 150 kW; the energy storage charge-discharge strategy specifies that within the next period of time, the charge-discharge of the energy storage system is reasonably arranged according to the energy supply and demand situation; the load priority allocation strategy is to preferentially shed a certain amount of commercial load, such as shedding 100 kW of commercial load, and appropriately reduce the amount of residential load shedding, such as shedding 100 kW of residential load, to ensure the stable operation of industrial loads.

[0098] Step S157: Repeatedly execute the optimized multi-source collaborative adjustment strategy set and iteratively update the parameters of the multi-source collaborative prediction model until the strategy execution deviation index is lower than the preset deviation threshold.

[0099] In this embodiment, after each execution of the multi-source collaborative adjustment strategy, new adjustment execution feedback data is obtained, and a new strategy execution deviation index is calculated. The new strategy execution deviation index is input again into the feedback learning layer of the multi-source collaborative prediction model to further adjust the feature fusion weight matrix of the encoder layer and the time window expansion parameter of the time series prediction layer. This process is continuously repeated until the strategy execution deviation index is lower than the preset deviation threshold. For example, the preset deviation threshold is 5. After multiple iterative adjustments, the strategy execution deviation index gradually decreases and finally drops to 4, meeting the requirement of being lower than the preset deviation threshold. At this time, under the action of the optimized strategy, the multi-source energy system can operate more stably and efficiently, achieving reasonable energy allocation and supply-demand balance.

[0100] In a possible implementation manner, step S155 includes:

[0101] Step S1551: Calculate the gradient change amount of the feature fusion weight matrix of the encoder layer according to the strategy execution deviation index, and update the feature fusion weight matrix based on the gradient descent algorithm.

[0102] In this embodiment, the strategy execution deviation index is obtained by comprehensively considering the deviations between the actual execution situations in multiple aspects and the prediction results. For example, after a certain execution of the multi-source collaborative adjustment strategy, the strategy execution deviation index is calculated. This strategy execution deviation index reflects the deviation situations of multiple factors on the energy supply side and the load demand side, including the differences between the actual values and the predicted values of photovoltaic power generation, wind power generation, charge-discharge of the energy storage system, industrial load, commercial load, and residential load, etc.

[0103] When calculating the gradient change of the feature fusion weight matrix in the encoder layer, it is necessary to analyze the contribution degree of each feature to the deviation. Taking the instantaneous power change rate of the photovoltaic power generation data sequence in the energy supply feature as an example, if there is a deviation between the actual change of the photovoltaic power generation power and the predicted situation, it is necessary to check whether the weight setting of this feature in the feature fusion weight matrix is reasonable. Suppose in the previous prediction, the weight of this feature was 0.3. By analyzing the policy execution deviation index, it is found that the prediction deviation of the photovoltaic power generation power has a certain impact on the overall deviation. During specific calculation, the actual change situation of the photovoltaic power generation power is compared with the predicted situation to obtain a deviation value related to this feature. For example, the actual change of the photovoltaic power generation power in a certain period is faster than the predicted value, resulting in a certain deviation. Through a complex calculation process (involving correlation analysis of this deviation value with the overall policy execution deviation index and considering various factors such as the role of this feature in the entire model), the gradient change of this feature weight is obtained as 0.05.

[0104] For the power standard deviation feature of the wind power generation data sequence, a similar analysis is also carried out. For example, suppose its original weight was 0.25. After comparing the actual fluctuation situation of the wind power generation power with the predicted result, the calculated gradient change of this feature weight is -0.03. This indicates that in the current situation, the weight of this feature in the fusion process may be too high and needs to be appropriately reduced.

[0105] For the charge-discharge switching frequency feature of the energy storage system output power data sequence, the original weight was 0.2. After analyzing the difference between the actual charge-discharge switching situation and the prediction, the calculated gradient change is 0.04. This means that the influence of this feature on the prediction result is greater than that reflected by the previously set weight, and the weight needs to be increased.

[0106] For the feature of the set of mutation time points of the industrial load demand data sequence in the load demand feature, the weight is 0.15. By comparing the difference between the actual industrial load mutation time and the predicted time, the calculated gradient change is 0.02.

[0107] For the load fluctuation amplitude feature of the commercial load demand data sequence, the weight is 0.1, and the calculated gradient change is -0.01.

[0108] For the periodic fluctuation pattern feature of the residential load demand data sequence, the weight is 0.1, and the calculated gradient change is 0.01.

[0109] For the cumulative offset amount feature of the power grid frequency deviation in the system stability deviation feature, the weight is 0.05, and the calculated gradient change is -0.005.

[0110] The instantaneous fluctuation extreme value feature of voltage deviation, with a weight of 0.05, and the calculated gradient change is 0.005.

[0111] Based on the gradient descent algorithm, the feature fusion weight matrix is updated. Taking the instantaneous power change rate feature weight of the photovoltaic power generation data sequence as an example, the updated weight is the original weight of 0.3 plus the gradient change of 0.05, that is, 0.3 + 0.05 = 0.35.

[0112] The power standard deviation feature weight of the wind power generation data sequence is updated to 0.25 + (-0.03) = 0.22.

[0113] The charge-discharge switching frequency feature weight of the energy storage system output power data sequence is updated to 0.2 + 0.04 = 0.24.

[0114] The mutation time point set feature weight of the industrial load demand data sequence is updated to 0.15 + 0.02 = 0.17.

[0115] The load fluctuation amplitude feature weight of the commercial load demand data sequence is updated to 0.1 + (-0.01) = 0.09.

[0116] The periodic fluctuation pattern feature weight of the residential load demand data sequence is updated to 0.1 + 0.01 = 0.11.

[0117] The cumulative offset feature weight of the power grid frequency deviation is updated to 0.05 + (-0.005) = 0.045.

[0118] The instantaneous fluctuation extreme value feature weight of voltage deviation is updated to 0.05 + 0.005 = 0.055.

[0119] In this way, the update of the feature fusion weight matrix of the encoder layer is completed, making the model more accurately reflect the actual situation when fusing various features.

[0120] Step S1552, execute the time dimension deviation component in the deviation index according to the strategy, and adjust the window length and sliding step in the time window expansion parameter of the time series prediction layer.

[0121] Among them, the adjustment of the time window expansion parameter satisfies the following conditions:

[0122] When the time dimension deviation component increases, shorten the window length and increase the sliding step.

[0123] When the time dimension deviation component decreases, extend the window length and decrease the sliding step.

[0124] In this embodiment, the time - dimension deviation component in the policy execution deviation index reflects the degree of deviation between the predicted time and the actual situation on the time axis.

[0125] Suppose that after a certain policy execution, the time - dimension deviation component in the calculated policy execution deviation index increases. This indicates that the accuracy of the predicted time has decreased, and it is necessary to adjust the time - window expansion parameters of the time - series prediction layer. For example, the current time - window length is 8 hours and the sliding step is 1 hour. Since the time - dimension deviation component has increased, shorten the window length and increase the sliding step according to the rules. Shorten the window length to 6 hours and increase the sliding step to 2 hours.

[0126] Specifically, when analyzing the time - series data of energy supply and load demand, it is found that when predicting with a window length of 8 hours and a sliding step of 1 hour before, many short - term change trends were not accurately captured, resulting in a large deviation between the predicted time and the actual situation. Shortening the window length to 6 hours can focus more on the short - term energy supply - demand changes and improve the prediction accuracy of the recent trends. Increasing the sliding step to 2 hours can, while ensuring a certain time coverage range, adapt to the data changes faster and adjust the prediction results in a timely manner.

[0127] Conversely, if the time - dimension deviation component in the policy execution deviation index decreases, it means that the accuracy of the predicted time has improved. For example, the original time - window length is 6 hours and the sliding step is 2 hours. At this time, since the time - dimension deviation component has decreased, extend the window length and reduce the sliding step. Extend the window length to 8 hours and reduce the sliding step to 1 hour. Such an adjustment is because in this case, a longer window length can better capture the trends of energy supply - demand over a longer time range, and a smaller sliding step can analyze the data changes more carefully, further improving the prediction accuracy.

[0128] By dynamically adjusting the window length and the sliding step in the time - window expansion parameters of the time - series prediction layer according to the time - dimension deviation component in the policy execution deviation index, the multi - source collaborative prediction model can better adapt to the dynamic changes of the multi - source energy system in the energy park, improve the prediction accuracy and timeliness, and ensure the stable and efficient operation of the energy system in the energy park. In the process of continuously repeating this process, the parameters of the model are continuously optimized, and the deviation between the prediction result and the actual situation is continuously reduced, ultimately achieving the precise regulation and optimal operation of the multi - source energy system.

[0129] In a possible implementation manner, before step S110, the method further includes a pre - training step of the multi - source collaborative prediction model:

[0130] Step S210, obtain a set of historical energy operation data, where the set of historical energy operation data includes historical energy supply characteristics, historical load demand characteristics, and historical system stability characteristics.

[0131] In this embodiment, in terms of historical energy supply characteristics, photovoltaic power generation data for the past year was collected. From sunrise to sunset every day, the photovoltaic power generation was recorded every 15 minutes. For example, on a certain day in spring, the photovoltaic power generation at 7:00 am was 100 kW, and it rose to 120 kW at 7:15 am. As the sunlight intensity increased, the power changed continuously, reaching a peak of 800 kW at 12:00 noon and then gradually decreasing. The photovoltaic power generation data for each day of the year was sorted in chronological order to form a record of photovoltaic power generation data. The same was true for wind power generation data, which fluctuated greatly under different seasons and weather conditions. For example, on a windy day in summer, the wind power generation increased rapidly from 500 kW to 700 kW within 1 hour. The long-term recorded wind power generation data was integrated. The output power data of the energy storage system recorded the power and time of each charge and discharge. For example, during a low electricity consumption period at night, the energy storage system was charged at a power of 150 kW for 3 hours. The above data was completely recorded to constitute the historical energy supply characteristics.

[0132] In terms of historical load demand characteristics, for industrial load data, there are many factories in the park. The production plans and equipment operation conditions of different factories are different, and the industrial load demand also changes accordingly. A large factory has a production period from 8:00 am to 5:00 pm on weekdays. The industrial load demand gradually rises from 1000 kW at 8:00 am to 1500 kW at 11:00 am, remains at a relatively high level from 2:00 pm to 4:00 pm, and then gradually decreases. The industrial load demand data of each factory in the past year was collected to form a record of industrial load data. In terms of commercial load data, the business hours of stores in the commercial area are relatively concentrated, such as from 9:00 am to 9:00 pm. The electricity consumption peaks are from 12:00 noon to 2:00 pm and from 6:00 pm to 8:00 pm. The load demand gradually rises from 300 kW at 9:00 am to 600 kW at 12:00 noon. The changes in commercial load throughout the year were recorded. Residential load data showed obvious daily periodicity. The electricity consumption peak was from 7:00 pm to 10:00 pm, rising from 400 kW at 7:00 pm to 600 kW at 9:00 pm. The electricity consumption trough was from 2:00 am to 5:00 am, remaining at about 150 kW. The above data was collected to form the historical load demand characteristics.

[0133] In terms of the historical system stability characteristics, the power grid frequency deviation data records the difference between the actual operating frequency and the standard frequency of the power grid every 30 minutes. For example, within a certain week, the recorded power grid frequency deviation data are -0.05 Hz, 0 Hz, 0.03 Hz, etc. The voltage deviation data, on the other hand, monitors the deviation of the power grid voltage from the rated voltage in real time. For example, within a certain month, the maximum recorded voltage deviation is +8 V, and the minimum is -5 V. The above data are organized to form the historical system stability characteristics.

[0134] Step S220: Perform standardization processing on the historical energy operation data set to generate a standardized training data set.

[0135] In a possible implementation manner, step S220 includes:

[0136] Step S221: Perform maximum-minimum normalization processing on the photovoltaic power generation data, wind power generation data, and energy storage system output power data in the historical energy supply characteristics to obtain the normalized energy supply characteristics.

[0137] Taking the photovoltaic power generation data as an example, among the data of the past year, the maximum power value is found to be 1000 kW, and the minimum power value is 50 kW. For the photovoltaic power generation value at a certain moment, such as 300 kW, perform normalization processing. The calculation process is as follows: Subtract the minimum value of 50 kW from this power value to get 300 - 50 = 250 kW, and then divide it by the difference between the maximum value and the minimum value, that is, 1000 - 50 = 950 kW, to obtain the normalized result of 250 ÷ 950 ≈ 0.263. Perform the same calculation on all photovoltaic power generation data to obtain the normalized photovoltaic power generation data. The wind power generation data and the energy storage system output power data are also normalized in this way, and finally the normalized energy supply characteristics are obtained.

[0138] Step S222: Perform Z-score standardization processing on the industrial load data, commercial load data, and residential load data in the historical load demand characteristics to obtain the standardized load demand characteristics.

[0139] Taking industrial load data as an example, first calculate the average value of the industrial load data. Assume that the total sum of industrial load data collected in a year is 5,475,000 kilowatts, and the number of records is 35,040 times (calculated based on 365 days in a year with 96 records per day), then the average value is 5,475,000÷35,040≈156.25 kilowatts. Then calculate the standard deviation. Square the difference between each industrial load data value and the average value, add up the above squared values, divide by the number of data, and finally take the square root. Assume that the standard deviation is approximately 200 kilowatts after calculation. For the industrial load value at a certain moment, such as 180 kilowatts, perform Z-score standardization processing. The calculation process is as follows: subtract the average value of 156.25 kilowatts from this value to get 180 - 156.25 = 23.75 kilowatts, and then divide by the standard deviation of 200 kilowatts, and the standardized result is 23.75÷200 = 0.119. Perform the same processing on all industrial load data to obtain the standardized industrial load data. The commercial load data and residential load data are also standardized in this way to obtain the standardized load demand characteristics.

[0140] Step S223: Perform a moving window mean process on the grid frequency deviation data and voltage deviation data in the historical system stability characteristics to obtain the smoothed system stability characteristics.

[0141] For the grid frequency deviation data, set the moving window size to 10 data points (i.e., 5 hours). For example, starting from the first data point, take the first 10 grid frequency deviation data as -0.05Hz, 0Hz, 0.03Hz, -0.02Hz, 0.01Hz, -0.01Hz, 0Hz, 0.02Hz, -0.03Hz, 0.04Hz. Add up the above data to get -0.05 + 0 + 0.03 + (-0.02) + 0.01 + (-0.01) + 0 + 0.02 + (-0.03) + 0.04 = 0.02Hz, and then divide by the number of data 10 to get the moving window mean of 0.02÷10 = 0.002Hz. Slide the window backward by one data point, take the second to the eleventh data points for the same calculation, and so on, to obtain a series of moving window means, and the above means constitute the smoothed grid frequency deviation data. For the voltage deviation data, also set the moving window size to 10 data points and calculate according to the above method to obtain the smoothed voltage deviation data, and then obtain the smoothed system stability characteristics.

[0142] Step S224: Align the normalized energy supply characteristics, the standardized load demand characteristics, and the smoothed system stability characteristics in time series and then merge them to generate a standardized training data set.

[0143] For example, the normalized photovoltaic power generation, wind power generation, and energy storage system output power data at the same time point are combined with the standardized industrial load, commercial load, and residential load data, as well as the smoothed grid frequency deviation and voltage deviation data to form a complete data record. All the above data records are sorted in chronological order and finally a standardized training data set is generated.

[0144] Step S230: Construct an initial multi-source collaborative prediction model, where the initial multi-source collaborative prediction model includes a feature encoder, a time series predictor, and a feedback corrector.

[0145] In this embodiment, the feature encoder is responsible for performing feature fusion on the input data and integrating different types of energy supply, load demand, and system stability features. The time series predictor performs time series analysis based on the fused features to predict future energy supply and load demand trends. The feedback corrector adjusts and optimizes the initial multi-source collaborative prediction model according to the difference between the prediction result and the real data.

[0146] Step S240: Input the standardized training data set into the initial multi-source collaborative prediction model for iterative training until the error rate between the prediction result output by the initial multi-source collaborative prediction model and the real historical operation data is lower than a preset threshold, and a pre-trained multi-source collaborative prediction model is obtained.

[0147] And step S240 includes:

[0148] Step S241: Divide the standardized training data set into a training set and a validation set, where the training set is used to update the parameters of the initial multi-source collaborative prediction model, and the validation set is used to evaluate the model generalization ability.

[0149] Assume it is divided according to a ratio of 7:3. Randomly select 70% of the data from the standardized training data set as the training set to update the parameters of the initial multi-source collaborative prediction model; the remaining 30% is used as the validation set to evaluate the generalization ability of the initial multi-source collaborative prediction model.

[0150] Step S242: In each iterative training process, call the feature encoder to perform feature fusion on the data in the training set to generate a training fusion feature vector.

[0151] For example, for a certain set of training data, the feature encoder fuses the normalized photovoltaic power generation data, wind power generation data, and energy storage system output power data with the standardized industrial load data, commercial load data, residential load data, as well as the smoothed power grid frequency deviation data and voltage deviation data. For example, weights can be assigned to each feature according to the importance of different features, and then the above features are combined according to the weights to generate a training fusion feature vector.

[0152] Step S243: Call the time series predictor to generate a training prediction result based on the training fusion feature vector.

[0153] In this embodiment, the time series predictor analyzes the time series information in the training fusion feature vector, combines the change trend of historical data, and predicts the energy supply and load demand in the future for a period of time. For example, it predicts the photovoltaic power generation, wind power generation, and dispatchable capacity of the energy storage system in the next 1 hour, as well as the industrial load, commercial load, and residential load demands.

[0154] Step S244: Call the feedback corrector to calculate the model loss function based on the difference between the training prediction result and the real historical operation data, and update the parameters of the feature encoder and the time series predictor based on the backpropagation algorithm.

[0155] Suppose the predicted photovoltaic power generation is 600 kW, while the photovoltaic power generation in the real historical operation data is 580 kW, and the difference between the two is 600 - 580 = 20 kW. The differences between all predicted values and real values are comprehensively calculated. For example, methods such as mean square error are used. After squaring the above difference values and summing them up and then averaging. For example, suppose after calculation, the total mean square error between all energy supply and load demand predicted values and real values is 100 (this is just an example to illustrate the calculation concept, and the actual calculation is more complex), and this value is the value of the model loss function.

[0156] Step S245: When the error rate on the validation set does not decrease continuously for multiple iterations, terminate the training and save the current model parameters as the pre-trained multi-source collaborative prediction model.

[0157] In this embodiment, the backpropagation algorithm calculates the contribution degree of each parameter to the loss from back to front according to the value of the model loss function, and then adjusts the parameters according to this contribution degree. For example, if it is found that the weight setting of a certain feature in the feature encoder causes a large prediction error, the weight will be adjusted according to the calculation result of the backpropagation algorithm. Suppose the original weight of a certain feature is 0.3, and after calculation and adjustment, the new weight may become 0.35. All parameters in the feature encoder and the time series predictor are adjusted in this way to reduce the value of the model loss function and improve the prediction accuracy of the model.

[0158] During the training process, the above steps are repeated continuously, and the model is optimized in each iteration. At the same time, the validation set is used to evaluate the generalization ability of the model. When the error rate on the validation set does not decrease for multiple consecutive iterations, it means that the model has reached a relatively stable state, and further training may not bring significant improvement. At this point, terminate the training and save the current model parameters as a pre-trained multi-source collaborative prediction model.

[0159] Figure 2 The schematic diagram shows exemplary hardware and software components of a multi-source energy debugging service system 100 that can implement the concept of the present application according to some embodiments of the present application. For example, the processor 120 can be used in the multi-source energy debugging service system 100 and used to perform the functions in the present application.

[0160] The multi-source energy debugging service system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the rebalancing dynamic adaptation method for a multi-source system of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0161] For example, the multi-source energy debugging service system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the multi-source energy debugging service system 100 may also include program instructions stored in ROM, RAM, or other types of non-temporary storage media, or any combination thereof. The method of the present application can be implemented according to the above program instructions. The multi-source energy debugging service system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0162] For ease of explanation, only one processor is described in the multi-source energy debugging service system 100. However, it should be noted that the multi-source energy debugging service system 100 in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the multi-source energy debugging service system 100 executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0163] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the rebalancing dynamic adjustment method applied to the multi-source system as described above is implemented.

[0164] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A rebalancing dynamic adjustment method applied to a multi-source system, characterized in that, The method includes: Collecting a real-time operation data set of a multi-source energy system, where the real-time operation data set includes energy supply characteristics, load demand characteristics, and system operation status characteristics; Performing dynamic supply-demand deviation feature extraction processing on the real-time operation data set to generate a dynamic supply-demand deviation feature set including supply fluctuation characteristics, demand fluctuation characteristics, and system stability deviation; Invoking a pre-trained multi-source collaborative prediction model to generate a multi-source collaborative prediction result based on the dynamic supply-demand deviation feature set, where the multi-source collaborative prediction result includes energy supply trend prediction characteristics and load demand trend prediction characteristics; Performing dynamic adaptation strategy matching processing based on the multi-source collaborative prediction result and the real-time operation data set to generate a multi-source collaborative adaptation strategy set, where the multi-source collaborative adaptation strategy set includes energy output adjustment strategies, energy storage charge and discharge strategies, and load priority allocation strategies; Executing the adaptation strategies in the multi-source collaborative adaptation strategy set and obtaining adaptation execution feedback data, and optimizing the parameters of the multi-source collaborative prediction model based on the adaptation execution feedback data to trigger a closed-loop adaptation operation; The performing dynamic supply-demand deviation feature extraction processing on the real-time operation data set to generate a dynamic supply-demand deviation feature set including supply fluctuation characteristics, demand fluctuation characteristics, and system stability deviation includes: Performing time series fluctuation analysis on the photovoltaic power data sequence, wind power data sequence, and energy storage system output power data sequence in the energy supply characteristics, extracting the instantaneous power change rate of the photovoltaic power data sequence as the first supply fluctuation sub-feature, extracting the power standard deviation of the wind power data sequence as the second supply fluctuation sub-feature, and extracting the charge and discharge switching frequency of the energy storage system output power data sequence as the third supply fluctuation sub-feature; Performing demand mutation detection processing on the industrial load demand data sequence, commercial load demand data sequence, and residential load demand data sequence in the load demand characteristics, determining the set of mutation time points of the industrial load demand data sequence as the first demand fluctuation sub-feature, determining the load fluctuation amplitude of the commercial load demand data sequence as the second demand fluctuation sub-feature, and determining the periodic fluctuation pattern of the residential load demand data sequence as the third demand fluctuation sub-feature; Performing stability index calculation on the grid frequency deviation data sequence and voltage deviation data sequence in the system operation status characteristics, obtaining the cumulative offset of the grid frequency deviation as the first stability deviation sub-feature, and obtaining the instantaneous fluctuation extreme value of the voltage deviation as the second stability deviation sub-feature; Combining the first supply fluctuation sub-feature, the second supply fluctuation sub-feature, the third supply fluctuation sub-feature, the first demand fluctuation sub-feature, the second demand fluctuation sub-feature, the third demand fluctuation sub-feature, the first stability deviation sub-feature, and the second stability deviation sub-feature to generate a dynamic supply-demand deviation feature set; The invoking a pre-trained multi-source collaborative prediction model to generate a multi-source collaborative prediction result based on the dynamic supply-demand deviation feature set includes: Input the dynamic supply-demand deviation feature set into the encoder layer of the multi-source collaborative prediction model. Through the encoder layer, perform multi-dimensional feature fusion processing on the dynamic supply-demand deviation feature set to generate a fusion feature vector containing supply-demand coupling features and system stability correlation features; Call the time series prediction layer of the multi-source collaborative prediction model to perform time series extension processing on the fusion feature vector, and generate energy supply trend prediction features and load demand trend prediction features within a future preset time window; Among them, the energy supply trend prediction features include a photovoltaic power generation power prediction curve, a wind power generation power prediction curve, and a schedulable capacity prediction curve of the energy storage system, and the load demand trend prediction features include an industrial load demand prediction curve, a commercial load demand prediction curve, and a residential load demand prediction curve; Calculate the supply-demand matching degree between the energy supply trend prediction features and the load demand trend prediction features to generate a multi-source collaborative prediction result including a supply-demand gap prediction value and a supply-demand balance time point; The execution of the adaptation strategies in the multi-source collaborative adaptation strategy set and obtaining the adaptation execution feedback data includes: Send the power adjustment instruction in the energy output adjustment strategy to the photovoltaic inverter controller of the multi-source energy system to obtain the actual adjustment response data of the photovoltaic power generation power; Send the charge-discharge rate instruction in the energy storage charge-discharge strategy to the bidirectional converter of the energy storage system to obtain the actual charge-discharge power data of the energy storage system; Send the load reduction instruction in the load priority allocation strategy to the load management terminal to obtain the actual load reduction amount data; Merge the actual adjustment response data, the actual charge-discharge power data, and the actual load reduction amount data to generate adaptation execution feedback data, and calculate the strategy execution deviation index between the adaptation execution feedback data and the multi-source collaborative prediction result; The parameter optimization of the multi-source collaborative prediction model based on the adaptation execution feedback data to trigger a closed-loop adaptation operation includes: Input the strategy execution deviation index into the feedback learning layer of the multi-source collaborative prediction model, and dynamically adjust the feature fusion weight matrix of the encoder layer and the time window extension parameters of the time series prediction layer through the feedback learning layer; Regenerate an updated multi-source collaborative prediction result according to the adjusted feature fusion weight matrix and time window extension parameters, and perform secondary dynamic adaptation strategy matching processing on the updated multi-source collaborative prediction result and the real-time operation data set to generate an optimized multi-source collaborative adaptation strategy set; Repeat the execution of the optimized multi-source collaborative adaptation strategy set and iteratively update the parameters of the multi-source collaborative prediction model until the strategy execution deviation index is lower than a preset deviation threshold.

2. The rebalancing dynamic adjustment method applied to a multi-source system according to claim 1, wherein The dynamic adaptation strategy matching process according to the multi-source collaborative prediction result and the real-time operation data set to generate a multi-source collaborative adaptation strategy set includes: Match the supply-demand gap prediction value with a preset multiple supply-demand balance threshold intervals to determine the target threshold interval to which the current supply-demand gap belongs; According to the policy generation rules corresponding to the target threshold interval, screen the set of schedulable energy supply sources from the energy supply characteristics, and determine the maximum charge-discharge power of the energy storage system based on the current state of charge and schedulable capacity in the energy storage charge-discharge strategy; According to the preset industry priority weights in the load priority allocation strategy, sort the industrial load demand data sequence, the commercial load demand data sequence, and the residential load demand data sequence in terms of load curtailment priority; Generate a multi-source collaborative adjustment strategy set including energy output adjustment strategy, energy storage charge-discharge strategy, and load priority allocation strategy based on the set of schedulable energy supply sources, the maximum charge-discharge power, and the load curtailment priority sorting; 3. The rebalancing dynamic adjustment method applied to a multi-source system according to claim 1, wherein The dynamic adjustment of the feature fusion weight matrix of the encoder layer and the time window expansion parameters of the time series prediction layer by the feedback learning layer includes: Calculate the gradient change of the feature fusion weight matrix of the encoder layer according to the policy execution deviation index, and update the feature fusion weight matrix based on the gradient descent algorithm; Adjust the window length and sliding step size in the time window expansion parameters of the time series prediction layer according to the time dimension deviation component in the policy execution deviation index; Among them, the adjustment of the time window expansion parameters satisfies the following conditions: When the time dimension deviation component increases, shorten the window length and increase the sliding step size; When the time dimension deviation component decreases, extend the window length and decrease the sliding step size.

4. The rebalancing dynamic adjustment method applied to a multi-source system according to claim 1, characterized in that, Before collecting the real-time operation data set of the multi-source energy system, the method further includes the pre-training step of the multi-source collaborative prediction model: Obtain the historical energy operation data set, which includes historical energy supply characteristics, historical load demand characteristics, and historical system stability characteristics; Perform standardization processing on the historical energy operation data set to generate a standardized training data set; Construct an initial multi-source collaborative prediction model, which includes a feature encoder, a time series predictor, and a feedback corrector; Input the standardized training data set into the initial multi-source collaborative prediction model for iterative training until the error rate between the prediction result output by the initial multi-source collaborative prediction model and the real historical operation data is lower than the preset threshold, and obtain the pre-trained multi-source collaborative prediction model.

5. The rebalancing dynamic adjustment method applied to a multi-source system according to claim 4, wherein The performing standardization processing on the historical energy operation data set to generate a standardized training data set includes: Perform maximum-minimum normalization processing on the photovoltaic power generation data, wind power generation data, and energy storage system output power data in the historical energy supply characteristics to obtain the normalized energy supply characteristics; Perform Z-score standardization processing on the industrial load data, commercial load data, and residential load data in the historical load demand characteristics to obtain the standardized load demand characteristics; Perform sliding window mean processing on the grid frequency deviation data and voltage deviation data in the historical system stability characteristics to obtain the smoothed system stability characteristics; Align the normalized energy supply characteristics, the standardized load demand characteristics, and the smoothed system stability characteristics in time series and then merge them to generate a standardized training data set; And, inputting the standardized training data set into the initial multi-source collaborative prediction model for iterative training until the error rate between the prediction result output by the initial multi-source collaborative prediction model and the real historical operation data is lower than a preset threshold to obtain a pre-trained multi-source collaborative prediction model, including: Dividing the standardized training data set into a training set and a validation set, wherein the training set is used to update the parameters of the initial multi-source collaborative prediction model, and the validation set is used to evaluate the model generalization ability; During each iterative training process, calling the feature encoder to perform feature fusion on the data in the training set to generate a training fusion feature vector; Calling the time series predictor to generate a training prediction result based on the training fusion feature vector; Calling the feedback corrector to calculate the model loss function according to the difference between the training prediction result and the real historical operation data, and updating the parameters of the feature encoder and the time series predictor based on the backpropagation algorithm; When the error rate on the validation set does not decrease for multiple consecutive iterations, terminate the training and save the current model parameters as the pre-trained multi-source collaborative prediction model.

6. A multi-source energy commissioning service system, characterized in that, The multi-source energy debugging service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the rebalancing dynamic adjustment method applied to the multi-source system according to any one of claims 1-5 above.

Images