Energy scheduling method and system

Abstract

The application discloses an energy scheduling method and system, which comprises the following steps: collecting operation data of a hydropower station in a target area, performing standardization processing and outlier elimination on the operation data, and establishing a hierarchical data structure DR including a real-time data layer DRT, a historical data layer DRH and a prediction data layer DRP based on the standardized data set; performing feature extraction and prediction modeling based on the hierarchical data structure DR, and constructing a real-time power calculation model, a preliminary total load prediction model and a final total load prediction model; and based on the prediction result of the final total load prediction model, adjusting the output power of adjustable units and the pumped storage power in real time, so that the difference between the real-time total load power and the prediction result of the final total load prediction model is within a preset range; the application organizes and classifies load prediction through a three-layer data structure, dynamically optimizes in combination with a multi-dimensional evaluation index system, and improves the scheduling precision and scheduling efficiency of the hydropower station.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydropower station dispatching, and in particular relates to an energy dispatching method and system. Background Art

[0002] The volatility of hydropower generation is an important factor affecting the stable operation of the power system. Accurately predicting load power and achieving reasonable scheduling are of great significance to improving the economy and reliability of hydropower stations.

[0003] Patent CN112930868A discloses a method for scheduling a hydropower station. This method relies on historical data to build a prediction model without considering the impact of real-time operating data, resulting in poor model adaptability and an inability to respond promptly to sudden changes in load. Furthermore, this method schedules all units uniformly without distinguishing between the characteristics of basic units and adjustable units, reducing scheduling flexibility. Patent CN114171431A proposes a method for hydropower station load prediction based on deep learning. Although this method introduces a neural network model to improve prediction accuracy, the data preprocessing process is overly simple and lacks a systematic outlier detection and processing mechanism, resulting in unstable input data quality and affecting the model's prediction effect. Furthermore, this method only focuses on short-term load prediction and does not fully consider the multi-timescale characteristics of load data, making it difficult to balance the impact of short-term fluctuations and long-term trends.

[0004] In summary, current hydropower station scheduling methods generally have problems such as imperfect data processing, poor adaptability of prediction models, and single scheduling strategies. There is an urgent need for a method that can effectively organize multi-time scale data, achieve refined load forecasting and flexible scheduling. Summary of the Invention

[0005] The purpose of the present invention is to provide an energy scheduling method and system. By constructing a three-layer data structure consisting of a real-time data layer (DRT), a historical data layer (DRH), and a predicted data layer (DRP), the effective organization and utilization of multi-time scale load data are achieved. By subdividing the output of basic units, adjustable unit output, and pumped storage power, a more refined load forecast model is implemented for optimized scheduling.

[0006] In a first aspect, the present invention provides an energy scheduling method for accurately predicting the load power of a hydropower station in a target area for the next six hours, and then scheduling the hydropower station based on the load power, the method comprising the following steps:

[0007] Step 1: collect the operating data of the hydropower stations in the target area, standardize the operating data and remove outliers to generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set.

[0008] The sampling period of the operation data of the collection target area hydropower station is set to 1 minute, and the hierarchical data structure DR includes a real-time data layer DRT, a historical data layer DRH, and a prediction data layer DRP. The data of each layer is a sequence with a length of 1440.

[0009] The real-time data layer DRT stores a standardized data set of the last 720 minutes, denoted as a DRT sequence, which is used to construct an actual power calculation model.

[0010] The data of the historical data layer DRH is converted from the DRT data, denoted as a DRH sequence, which is used to construct a preliminary total load prediction model.

[0011] The data of the prediction data layer DRP is converted from the DRH data, denoted as a DRP sequence.

[0012] Step two, feature extraction and prediction modeling based on the hierarchical data structure DR, using DRT data to construct a real-time power calculation model, based on DRH data and real-time power calculation model to construct a preliminary total load prediction model; combining DRT and DRH data to construct a final total load prediction model.

[0013] Step three, based on the prediction result of the final total load prediction model, real-time adjustment of the output power of the adjustable unit and the pumped storage power, so that the difference between the real-time total load power and the prediction result of the final total load prediction model is within a preset range, and the energy scheduling is completed.

[0014] Further, the method for standardizing the operation data in step one is: Where X Rt is the original data, X Rt , max and X Rt , min are the maximum and minimum values of the data, respectively, and X ′ Rt is the standardized DRT data;

[0015] Anomaly detection is performed using a sliding time window |X ′ Rt -μ w |>3σ w , where μ w and σ w are the mean and standard deviation of the data in the time window, respectively, and the window length is 12 hours;

[0016] The standardization method for DRH data is: Where X RH is the DRH data before standardization, μ RH and σ RHis the mean and standard deviation of DRH data, X ′ RH is the standardized DRH data; the standardization method for DRP data is: Among them, X RP is the DRP data before standardization, X RH , min and X RH , max are the minimum and maximum values ​​of the standardized DRH data;

[0017] X RH，new =0.8X Rt，avg +0.2X RH，old ;

[0018] The mapping from DRH data to DRP data is expressed as: RP =X RH +0.1ΔX trend +ε;

[0019] Among them, X RH , new and X RH , old Represents the new DRH data and the old DRH data respectively, X Rt , avg represents the mean of DRT data; ΔX trend It represents the average value of the difference between the new DRH data and the old DRH data; ε is the trend term coefficient, which is determined based on the historical data fitting error.

[0020] Furthermore, using DRT data to build a short-term prediction model includes: establishing a real-time total load power L based on DRT data total (t) Prediction model:

[0021] L total (t) = L1(t) + L2(t) + L3(t); where L total (t) is the real-time total load power at time t, where t is in minutes. L1(t) is the basic unit output power at time t, which refers to the minimum load required to maintain the basic operation of the power station. L2(t) is the adjustable unit output power at time t, which allows flexible scheduling within [t, t+360] minutes. L3(t) is the pumped storage power at time t, with the maximum power not exceeding 40% of the rated power.

[0022] The preliminary total load forecasting model based on DRH data is:

[0023]

[0024] in, The total load power is The preliminary prediction results of is the prediction duration, P base Indicates the minimum power demand in the last 360 minutes. P peak is the peak power amplitude, which indicates the load fluctuation amplitude in the last 360 minutes. is the phase angle, reflecting the load peak and valley time distribution; for The random fluctuation component of the load.

[0025] Combine DRT and DRH data to generate the final total load forecast model:

[0026] in, The total load power is The final prediction result at for Load change trend item at time for The prediction error term when , satisfies the normal distribution N(0, σ 2 ).

[0027] Furthermore, the method for adjusting the output power of the adjustable unit and the pumped storage power in real time in step 3 includes:

[0028] when When the adjustable unit output L2(t) is adjusted first, The load power is evenly distributed within the system; when the output of the adjustable unit still does not meet the requirements after adjustment, the pumped storage power L3(t) is reduced, and the reduction rate does not exceed 30%.

[0029] Real-time monitoring of the adjusted total load power L total (t), ensure L total (t) and P pred The deviation of (t) does not exceed 5%.

[0030] when When the output of the adjustable unit is adjusted, the distribution ratio of L2(t) is increased. If there is still a gap after the output of the adjustable unit is adjusted, the power of the pumped storage power L3(t) is increased, but it must be ensured that it does not exceed 120% of the rated power of the pumped storage power.

[0031] Furthermore, a prediction model evaluation index system is established:

[0032] Mean absolute error

[0033] Root mean square error

[0034] Prediction accuracy

[0035] Dynamically optimize the prediction model parameters based on the evaluation results of the evaluation index system: when MAE>10%, recalculate the trend term coefficient ε; when RMSE>15%, update P var (t) calculation weight, when ACC < 85%, adjust P trend (t) is the prediction period.

[0036] Furthermore, the random load fluctuation component P var The calculation method of (t) is:

[0037] Where ΔP avg ΔP is the average deviation of the load power in the same period within the last 24 hours; std is the standard deviation of the load power in the last 24 hours; when |ΔP avg |<ΔP std When P var(t) =ΔP std .

[0038] Furthermore, the load change trend item The calculation method is:

[0039] Among them, P now is the load power P at the current time t past For the moment Load power in the previous 24 hours;

[0040] When the calculated When it is greater than the historical maximum change rate, the historical maximum change rate value is taken.

[0041] Furthermore, the prediction error term P err The calculation method of (t) is: Among them, P max_err is the maximum prediction error in the last 7 days, τ is the characteristic time constant, and its value is 12; when minute, The values ​​are:

[0042] Where α is the attenuation coefficient, and its value range is [0.01, 0.05].

[0043] Furthermore, when the MAE is greater than 12% in three consecutive evaluations, the model retraining mechanism is triggered;

[0044] When RMSE>15% and ACC<85%, data quality assessment was re-performed and abnormal data points were removed;

[0045] Conduct a complete model evaluation every 24 hours, and adaptively adjust the prediction period and model parameters based on the evaluation results;

[0046] The model optimization adopts a sliding time window mechanism with a window length of 10080 minutes and a step size of 10080 minutes.

[0047] In the second aspect, based on the same inventive concept, the present invention provides an energy scheduling system for executing the energy scheduling method described in the first aspect, wherein the system includes: a data acquisition and processing module, a prediction modeling module, and an energy scheduling module connected in sequence.

[0048] The data acquisition and processing module is used to collect the operating data of the hydropower stations in the target area, perform standardization processing on the operating data and eliminate outliers, generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set.

[0049] The sampling period of the data acquisition and processing module is set to 1 minute. The hierarchical data structure DR includes: a real-time data layer DRT, a historical data layer DRH and a predicted data layer DRP. The data of each layer is composed of a sequence with a length of 1440.

[0050] The real-time data layer DRT stores the standardized data set of the latest 12 hours, which is recorded as a DRT sequence and is used to construct an actual power calculation model.

[0051] The data of the historical data layer DRH is aggregated and transformed from the DRT data, recorded as a DRH sequence, and used to construct a preliminary total load prediction model.

[0052] The data of the prediction data layer DRP is obtained by converting the DRH data and is recorded as a DRP sequence.

[0053] The prediction modeling module is used to perform feature extraction and prediction modeling based on the hierarchical data structure DR, use DRT data to build a real-time power calculation model, build a preliminary total load prediction model based on DRH data and the real-time power calculation model; and build a final total load prediction model by combining DRT and DRH data.

[0054] The energy scheduling module is used to adjust the output power of the adjustable units and the pumped storage power in real time based on the prediction results of the final total load prediction model, so that the difference between the real-time total load power and the prediction results of the final total load prediction model is within a preset range, thereby completing energy scheduling.

[0055] The beneficial effects of the present invention are as follows:

[0056] The present invention realizes the effective organization and utilization of multi-time-scale load data by constructing a three-layer data structure consisting of a real-time data layer DRT, a historical data layer DRH, and a prediction data layer DRP. By subdividing the load into basic unit output, adjustable unit output, and pumped storage power, a more refined load prediction model is established by combining a real-time power calculation model with a preliminary total load prediction model. The prediction model parameters are dynamically optimized based on the three evaluation indicators of MAE, RMSE, and ACC, and the model is optimized using a sliding time window mechanism, which significantly improves the prediction accuracy. More efficient energy scheduling and control are achieved through the flexible scheduling of adjustable unit output and the reasonable reduction of pumped storage power. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of an energy scheduling method of the present invention;

[0058] Figure 2 This is a schematic diagram of the energy scheduling system of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0060] Example 1

[0061] like Figure 1 As shown, an energy scheduling method of the present invention is used to accurately predict the load power of a hydropower station in a target area in the next 6 hours, and then schedule the hydropower station based on the load power. The method includes the following steps:

[0062] Step 1: collect the operating data of the hydropower stations in the target area, standardize the operating data and remove outliers to generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set.

[0063] For the operation data of a hydropower station, usually includes the following core aspects: real-time power output of the generator unit, reservoir water level, inflow, outflow, unit speed, guide vane opening of the water turbine, tail water level and other basic operation parameters, and power generation efficiency of the hydropower station, unit vibration, bearing temperature, stator temperature and other equipment state parameters; in addition, the dispatching model in the prior art also needs to collect related environmental data such as upstream rainfall, air temperature, weather conditions and the like. The present application converts the above operation data into real-time data at the power level, and only focuses on the power of each unit or load, thereby simplifying the dispatching operation; the operation data of the present application includes: real-time total load power, basic unit output power, adjustable unit output power and pumped storage power.

[0064] The sampling period of the operation data of the collection target area hydropower station is set to 1 minute, and the hierarchical data structure DR includes: a real-time data layer DRT, a historical data layer DRH and a prediction data layer DRP, and the data of each layer is a sequence with a length of 1440.

[0065] The three-layer data structure design adopted by the present application fully considers the time characteristics of the power load data. The 720-minute data length selection of the real-time data layer DRT is based on the intraday fluctuation period of the load, which can effectively capture the short-term load change characteristics; the historical data layer DRH realizes the extraction of long-term trends through data aggregation, and the prediction data layer DRP establishes a bridge between the historical mode and the future prediction through the mapping relationship. This hierarchical data organization method not only ensures real-time performance, but also takes into account the use of historical rules.

[0066] The real-time data layer DRT saves the latest 720-minute standardized data set, denoted as DRT sequence, for constructing an actual power calculation model.

[0067] The data of the historical data layer DRH is converted from the DRT data aggregation, denoted as DRH sequence, for constructing a preliminary total load prediction model.

[0068] The data of the prediction data layer DRP is converted from the DRH data, denoted as DRP sequence.

[0069] Step two, feature extraction and prediction modeling based on the hierarchical data structure DR, using DRT data to construct a real-time power calculation model, based on DRH data and real-time power calculation model to construct a preliminary total load prediction model; combining DRT and DRH data to construct a final total load prediction model.

[0070] Step three, based on the prediction result of the final total load prediction model, real-time adjusting the adjustable unit output power and the pumped storage power, so that the difference between the real-time total load power and the prediction result of the final total load prediction model is within a preset range, completing energy dispatching.

[0071] The method for standardizing the operating data in step 1 is: Among them, X Rt is the original data, X Rt，max and X Rt，min are the maximum and minimum values ​​of the data, respectively, X′ Rt is the standardized DRT data;

[0072] Outlier detection using sliding time window |X′ Rt -μ w |>3σ w , where μ w and σ w are the mean and standard deviation of the data in the time window, respectively, and the window length is 12 hours;

[0073] The standardized processing method for DRH data is: Among them, X RH is the DRH data before normalization, μ RH and σ RH is the mean and standard deviation of DRH data, X′ RH is the standardized DRH data; the standardization method for DRP data is: Among them, X RP is the DRP data before standardization, X RH，min and X RH，max are the minimum and maximum values ​​of the standardized DRH data;

[0074] X RH，new =0.8X Rt，avg +0.2X RH，old ;

[0075] The mapping from DRH data to DRP data is expressed as: RP =X RH +0.1ΔX trend +ε;

[0076] Among them, X RH，new and X RH，old Represents the new DRH data and the old DRH data respectively, X Rt，avg represents the mean of DRT data; ΔX trend It represents the average value of the difference between the new DRH data and the old DRH data; ε is the trend term coefficient, which is determined based on the historical data fitting error.

[0077] For DRT data, the use of maximum and minimum value standardization can maintain the relative change relationship of the data; the use of Z-score standardization for DRH data highlights the statistical characteristics of the data; and the standardization of DRP data based on the value range of DRH data achieves the scale consistency of predicted data and historical data; the combined application of the three standardization methods better adapts to the characteristics of different data levels.

[0078] Using DRT data to build a short-term forecast model includes: building a real-time total load power L based on DRT data total (t) Prediction model:

[0079] L total (t) = L1(t) + L2(t) + L3(t); where L total (t) is the real-time total load power at time t, where t is in minutes. L1(t) is the basic unit output power at time t, which refers to the minimum load required to maintain the basic operation of the power station. L2(t) is the adjustable unit output power at time t, which allows flexible scheduling within [t, t+360] minutes. L3(t) is the pumped storage power at time t, with the maximum power not exceeding 40% of the rated power.

[0080] The preliminary total load forecasting model based on DRH data is:

[0081]

[0082] in, The total load power is The preliminary prediction results of is the prediction duration, P base Indicates the minimum power demand in the last 360 minutes. P peak is the peak power amplitude, which indicates the load fluctuation amplitude in the last 360 minutes. is the phase angle, reflecting the load peak and valley time distribution; for The random fluctuation component of the load.

[0083] Combine DRT and DRH data to generate the final total load forecast model:

[0084] in, The total load power is The final prediction result at for Load change trend item at time for The prediction error term when , satisfies the normal distribution N(0, σ 2 ).

[0085] The load forecasting model design embodies the principle of "layered forecasting and gradual optimization." The base load, Pbase, reflects stable electricity demand, the sinusoidal term characterizes typical cyclical fluctuations, and the random fluctuation component, Pvar, describes short-term disturbances. This decomposed modeling approach ensures the model's ability to represent key load characteristics while providing the flexibility to optimize different components individually.

[0086] The method for real-time adjustment of the output power of the adjustable unit and the pumped storage power in step 3 includes:

[0087] when When the adjustable unit output L2(t) is adjusted first, The load power is evenly distributed within the system; when the output of the adjustable unit still does not meet the requirements after adjustment, the pumped storage power L3(t) is reduced, and the reduction rate does not exceed 30%.

[0088] Real-time monitoring of the adjusted total load power L total (t), ensure L total (t) and P pred The deviation of (t) does not exceed 5%.

[0089] when When the output of the adjustable unit is adjusted, the distribution ratio of L2(t) is increased. If there is still a gap after the output of the adjustable unit is adjusted, the power of the pumped storage power L3(t) is increased, but it must be ensured that it does not exceed 120% of the rated power of the pumped storage power.

[0090] Establish a prediction model evaluation indicator system:

[0091] Mean absolute error

[0092] Root mean square error

[0093] Prediction accuracy

[0094] Dynamically optimize the prediction model parameters based on the evaluation results of the evaluation index system: when MAE>10%, recalculate the trend term coefficient ε; when RMSE>15%, update P var (t) calculation weight, when ACC < 85%, adjust P trend (t) is the prediction period.

[0095] MAE measures the overall prediction bias, RMSE is more sensitive to large deviations, and ACC provides a perspective of relative error. Based on the differences in different indicators, the optimization strategy is set, such as when MAE is large, the trend item is adjusted first, and when RMSE is large, the random fluctuation component is focused on. This fine-tuned optimization mechanism significantly improves the adaptive ability of the model.

[0096] The load random fluctuation component P var The calculation method of (t) is:

[0097] Where ΔP avg is the average deviation of the load power in the same period within the last 24 hours; ΔP std is the standard deviation of the load power in the last 24 hours; when |ΔP avg |<ΔP std , take P var(t) =ΔP std .

[0098] The load change trend item P The calculation method is:

[0099] Where P now is the load power P past at the current time t; P err is the load power at time 24 hours ago.

[0100] When the calculated is greater than the historical maximum change rate, take the historical maximum change rate value.

[0101] The calculation method of the prediction error item P err (t) is: Where P max_err is the maximum prediction error in the last 7 days, and τ is the characteristic time constant, which is 12; when minutes, the value is:

[0102] Where α is the decay coefficient, and the value range is [0.01, 0.05].

[0103] When MAE is greater than 12% in three consecutive evaluations, trigger the model retraining mechanism;

[0104] When RMSE>15% and ACC<85%, re-evaluate the data quality and remove abnormal data points;

[0105] The model is evaluated once every 24 hours, and the prediction period and model parameters are adjusted adaptively according to the evaluation results;

[0106] The model optimization adopts a sliding time window mechanism with a window length of 10080 minutes and a step size of 10080 minutes.

[0107] Example 2

[0108] like Figure 2 FIG. 1 is a schematic diagram of the composition of an energy scheduling system of the present invention, which is used to execute the energy scheduling method of Example 1.

[0109] The system includes: a data acquisition and processing module, a prediction modeling module and an energy scheduling module which are connected in sequence.

[0110] The system architecture adopts a modular design, with standardized interfaces enabling data exchange between functional modules. The data acquisition and processing module not only performs basic data preprocessing but also assumes the important responsibility of data quality control. The predictive modeling module adopts a progressive model construction strategy to ensure the reliability of prediction results. The energy scheduling module implements closed-loop control based on the prediction results, ensuring stable system operation through real-time monitoring and dynamic adjustments. This layered and progressive system design effectively improves the reliability and maintainability of the entire system.

[0111] The data acquisition and processing module is used to collect the operating data of the hydropower stations in the target area, perform standardization processing on the operating data and eliminate outliers, generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set.

[0112] The sampling period of the data acquisition and processing module is set to 1 minute. The hierarchical data structure DR includes: a real-time data layer DRT, a historical data layer DRH and a predicted data layer DRP. The data of each layer is composed of a sequence with a length of 1440.

[0113] The real-time data layer DRT stores the standardized data set of the latest 12 hours, which is recorded as a DRT sequence and is used to construct an actual power calculation model.

[0114] The data of the historical data layer DRH is aggregated and transformed from the DRT data, recorded as a DRH sequence, and used to construct a preliminary total load prediction model.

[0115] The data of the prediction data layer DRP is obtained by converting the DRH data and is recorded as a DRP sequence.

[0116] The prediction modeling module is used to perform feature extraction and prediction modeling based on the hierarchical data structure DR, use DRT data to build a real-time power calculation model, build a preliminary total load prediction model based on DRH data and the real-time power calculation model; and build a final total load prediction model by combining DRT and DRH data.

[0117] The energy scheduling module is used to adjust the output power of the adjustable units and the pumped storage power in real time based on the prediction results of the final total load prediction model, so that the difference between the real-time total load power and the prediction results of the final total load prediction model is within a preset range, thereby completing energy scheduling. The preset range is generally set to a difference of no more than 2% between the two values.

[0118] When the predicted load deviates from the actual load, the fluctuation is first smoothed by adjusting the output of the adjustable units. This "flexible regulation priority" strategy reduces frequent intervention in the pumped storage power. Furthermore, by setting a reasonable power regulation range (maximum pumped storage power reduction rate of 30%), regulation margin is ensured while avoiding system instability caused by over-regulation.

[0119] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An energy dispatching method for accurately predicting the load power of a hydropower station in a target area for the next 6 hours, and then dispatching the hydropower station based on the load power, characterized in that: The method comprises the following steps: Step 1: collect the operating data of the hydropower stations in the target area, standardize the operating data and remove outliers to generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set; The operating data of the hydropower station in the target area includes: real-time total load power, basic unit output power, adjustable unit output power and pumped storage power; The sampling period for collecting the operating data of the hydropower stations in the target area is set to 1 minute. The hierarchical data structure DR includes: a real-time data layer DRT, a historical data layer DRH and a prediction data layer DRP. The data of each layer is composed of a sequence with a length of 1440; The real-time data layer DRT stores the standardized data set of the last 720 minutes, which is recorded as a DRT sequence and used to build the actual power calculation model; The data of the historical data layer DRH is aggregated and converted from DRT data and recorded as DRH sequence, which is used to build a preliminary total load prediction model; The data of the prediction data layer DRP is obtained by converting the DRH data and is recorded as a DRP sequence; Step 2: Perform feature extraction and predictive modeling based on the hierarchical data structure DR, use DRT data to build a real-time power calculation model, and build a preliminary total load prediction model based on DRH data and the real-time power calculation model; and build a final total load prediction model by combining DRT and DRH data. Using DRT data to build a short-term forecast model includes: establishing real-time total load power based on DRT data Prediction Model: ;in, is the real-time total load power at time t, where t is in minutes. is the basic unit output power at time t, which refers to the minimum load to maintain the basic operation of the power station; is the adjustable unit output power at time t, which allows flexible scheduling within [t, t+360] minutes. is the pumped storage power at time t, and the maximum power does not exceed 40% of the rated power; The preliminary total load forecasting model based on DRH data is: ； in, The total load power is The preliminary prediction results of To predict the duration, Indicates the minimum power demand in the last 360 minutes. ; is the peak power amplitude, which indicates the load fluctuation amplitude in the last 360 minutes. is the phase angle, reflecting the load peak and valley time distribution; for The random fluctuation component of the load when ; Combine DRT and DRH data to generate the final total load forecast model: ;in, The total load power is The final prediction result at for Load change trend item at time for The prediction error term when , satisfies the normal distribution ; Step three: Based on the prediction results of the final total load prediction model, the output power of the adjustable units and the pumped storage power are adjusted in real time so that the difference between the real-time total load power and the prediction results of the final total load prediction model is within the preset range, thereby completing energy scheduling.

2. The energy scheduling method according to claim 1, characterized in that: The method for standardizing the operating data in step 1 is: ,in, is the original data, and are the maximum and minimum values ​​of the data, respectively. is the standardized DRT data; Outlier detection using sliding time windows ,in and are the mean and standard deviation of the data in the time window, respectively, and the window length is 12 hours; The standardized processing method for DRH data is: ;in, is the DRH data before normalization, and is the mean and standard deviation of DRH data, is the standardized DRH data; the standardization method for DRP data is: ,in, is the DRP data before standardization, and is the minimum and maximum value of the standardized DRH data; ； The mapping from DRH data to DRP data is expressed as: ; in, and Represents the new DRH data and the old DRH data respectively, represents the mean of DRT data; Represents the average value of the difference between the new DRH data and the old DRH data; is the trend term coefficient, which is determined based on the fitting error of historical data.

3. The energy scheduling method according to claim 2, characterized in that: The method for real-time adjustment of the output power of the adjustable unit and the pumped storage power in step 3 includes: when When adjusting the output of the adjustable unit, first adjust the output of the adjustable unit. ,exist Evenly distribute the load power; when the adjustable unit output still does not meet the requirements after adjustment, reduce the pumped storage power , the reduction rate shall not exceed 30%; Real-time monitoring of the total load power after adjustment ,make sure and The deviation does not exceed 5%; when When the adjustable unit output is increased If there is still a gap after the adjustable unit output is adjusted, increase the pumped storage power power, but it must not exceed 120% of the rated power of the pumped storage power.

4. The energy scheduling method according to claim 3, characterized in that: Establish a prediction model evaluation indicator system: Mean absolute error ; Root mean square error ; Prediction accuracy ; Dynamically optimize the prediction model parameters based on the evaluation results of the evaluation index system: when MAE>10%, recalculate the trend term coefficient , when RMSE>15%, update The calculation weight of ACC is adjusted when ACC < 85%. prediction period.

5. The energy scheduling method according to claim 4, characterized in that: The random load fluctuation component The calculation method is: ;in, The average deviation of the load power in the same period in the last 24 hours; is the standard deviation of load power in the last 24 hours; When, take .

6. The energy scheduling method according to claim 5, characterized in that: The load change trend item The calculation method is: ;in, is the load power at the current time t For the moment Load power in the previous 24 hours; When the calculated When it is greater than the historical maximum change rate, the historical maximum change rate value is taken.

7. The energy scheduling method according to claim 6, characterized in that: The prediction error term The calculation method is: ,in, is the maximum forecast error in the last 7 days, is the characteristic time constant, which takes a value of 12; when minute, The values ​​are: ,in is the attenuation coefficient, and its value range is [0.01, 0.05].

8. The energy scheduling method according to claim 7, characterized in that: When the MAE is greater than 12% in three consecutive evaluations, the model retraining mechanism is triggered; When RMSE>15% and ACC<85%, data quality was reassessed and abnormal data points were removed; Conduct a complete model evaluation every 24 hours, and adaptively adjust the prediction period and model parameters based on the evaluation results; The model optimization adopts a sliding time window mechanism with a window length of 10080 minutes and a step size of 10080 minutes.

9. An energy scheduling system, used to execute the energy scheduling method according to any one of claims 1 to 8, characterized in that: The system includes: a data acquisition and processing module, a prediction modeling module, and an energy scheduling module connected in sequence; The data acquisition and processing module is used to collect the operating data of the hydropower stations in the target area, perform standardization processing on the operating data and eliminate outliers to generate a standardized data set, and establish a hierarchical data structure DR based on the standardized data set; The sampling period of the data acquisition and processing module is set to 1 minute. The hierarchical data structure DR includes: a real-time data layer DRT, a historical data layer DRH and a predicted data layer DRP. The data of each layer is composed of a sequence with a length of 1440; The real-time data layer DRT stores the standardized data set of the last 720 minutes, which is recorded as a DRT sequence and used to build the actual power calculation model; The data of the historical data layer DRH is aggregated and converted from DRT data and recorded as DRH sequence, which is used to build a preliminary total load prediction model; The data of the prediction data layer DRP is obtained by converting the DRH data and is recorded as a DRP sequence; The prediction modeling module is used to perform feature extraction and prediction modeling based on the hierarchical data structure DR, build a real-time power calculation model using DRT data, build a preliminary total load prediction model based on DRH data and the real-time power calculation model; and build a final total load prediction model by combining DRT and DRH data; The energy scheduling module is used to adjust the output power of the adjustable units and the pumped storage power in real time based on the prediction results of the final total load prediction model, so that the difference between the real-time total load power and the prediction results of the final total load prediction model is within a preset range, thereby completing energy scheduling.

Images