Multi-source climbing collaborative prediction system containing pumped storage power system and modeling method

Abstract

The invention relates to the technical field of power system automation, in particular to a multi-source climbing collaborative prediction system containing a pumped storage power system and a modeling method, and the system comprises a component decomposition module which is used for decomposing the total climbing demand of the system into a load climbing component, a wind power climbing component and a photovoltaic climbing component based on the physical mechanism of the power system; the differential parallel prediction module comprises a first LSTM branch network used for processing the load climbing component, a second LSTM branch network used for processing the wind power climbing component and a third LSTM branch network used for processing the photovoltaic climbing component, and the context sensing dynamic fusion module is used for receiving output characteristics of the LSTM branch networks and outputting the output characteristics of the LSTM branch networks. And based on the context vector containing the system operation state, dynamically calculating the weight of each component through an attention mechanism, and carrying out weighted fusion to generate a final system total climbing demand prediction value. According to the method, calculation redundancy is reduced through specialized design, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a multi-source ramp-up collaborative prediction system and modeling method for pumped storage power systems. Background Technology

[0002] Existing power system ramp demand forecasting technologies, including pumped storage systems, primarily rely on traditional time series forecasting methods and deep learning approaches to predict the system's net load change rate. Commonly used techniques include ARIMA models, support vector regression (SVR), and neural networks for ramp demand forecasting. These methods offer advantages such as relatively mature algorithms, ease of implementation, and the ability to provide a certain level of accuracy under stable operating conditions. However, these traditional forecasting methods have limitations when dealing with complex power systems with high renewable energy penetration rates. Particularly in situations involving severe coupling of multi-source uncertainties or rapid changes in system operating conditions, they may experience decreased prediction accuracy, insufficient interpretability, or low computational efficiency.

[0003] Existing technical solutions mainly rely on holistic black-box modeling methods, which are insufficient in terms of multi-source uncertainty separation and physical mechanism fusion. Traditional prediction algorithms often require a large amount of historical data for training and have a relatively slow response speed, making them unable to effectively cope with the complex prediction needs in the context of high penetration of new energy sources. In particular, for the prediction of power system ramp-up demand, existing prediction technologies are unable to provide interpretable analysis of the contributions of each source in dynamic changes, which can easily lead to difficulties in scheduling decisions or increased system operation risks. Summary of the Invention

[0004] To address this, the present invention provides a multi-source ramp-up collaborative prediction system for pumped storage power systems, which overcomes the problem in the prior art that it is difficult to provide interpretable analysis of the contributions of each source in dynamic changes, thus leading to difficulties in scheduling decisions or increased risks in system operation.

[0005] To achieve the above objectives, the present invention provides a multi-source ramp-up collaborative prediction system including a pumped-storage power system, comprising:

[0006] The component decomposition module is used to decompose the total ramp demand of the power system into load ramp components, wind power ramp components, and photovoltaic ramp components based on the physical mechanism of the power system.

[0007] The differentiated parallel prediction module, connected to the component decomposition module, includes a first LSTM branch network for processing the load ramping component, a second LSTM branch network for processing the wind power ramping component, and a third LSTM branch network for processing the photovoltaic ramping component. The first LSTM branch network, the second LSTM branch network, and the third LSTM branch network adopt differentiated network architectures based on the time-series characteristics of their respective components.

[0008] The context-aware dynamic fusion module, connected to the differentiated parallel prediction module, is used to receive the output features of each LSTM branch network, and based on the context vector containing the system operating state, dynamically calculate the weight of each component through an attention mechanism, perform weighted fusion, and generate the final predicted value of the total system ramp demand.

[0009] Furthermore, the first LSTM branch network, the second LSTM branch network, and the third LSTM branch network adopt differentiated network architectures based on the temporal characteristics of their respective components.

[0010] The first LSTM branch network is a three-layer LSTM structure and uses the tanh activation function;

[0011] The second LSTM branch network is a two-layer LSTM structure, and uses the ReLU activation function and layer normalization;

[0012] The third LSTM branch network is a two-layer LSTM structure, and it adopts the Swish activation function and gated linear unit mechanism.

[0013] Furthermore, the context-aware dynamic fusion module adopts a multi-head attention mechanism, and the context vector includes one or more of the following: new energy penetration rate, system load rate, weather type, and seasonal information.

[0014] Furthermore, the multi-source hill-climbing collaborative prediction system also includes a unified training and optimization framework, which trains the system using a multi-objective loss function, including:

[0015] Total system ramp demand prediction loss;

[0016] The weighted sum of the predicted losses for each component;

[0017] The physical consistency constraint is used to ensure consistency between the sum of the predicted values ​​of each component and the predicted value of the total ramp demand of the system.

[0018] A modeling method for a multi-source ramp-up collaborative prediction system applied to the pumped-storage power system includes:

[0019] Step S1: Collect multi-source operation data of the power system and decompose the total ramp demand of the system into load ramp component, wind power ramp component and photovoltaic ramp component based on physical mechanism;

[0020] Step S2: Perform specialized feature engineering on each component to construct the corresponding input features;

[0021] Step S3: Input the processed component data into the corresponding differential LSTM branch network and calculate the predicted feature vector of each component in parallel.

[0022] Step S4: Construct the system running context vector, dynamically calculate the fusion weight of the predicted features of each component using the attention mechanism, perform weighted fusion, and output the final prediction result of the total system ramp demand.

[0023] Step S5: Based on the prediction error feedback, update the network parameters using a multi-objective loss function.

[0024] Furthermore, in step S1, the total system ramp demand is decomposed into load ramp component, wind power ramp component, and photovoltaic ramp component based on physical mechanisms, including:

[0025] The system's ramp-up requirement decomposition formula is as follows:

[0026] ;

[0027] The physical modeling methods for each component are as follows:

[0028] Load ramp component: Modeling is based on the periodic characteristics of the load and the influence of temperature;

[0029] In the formula, This represents the load ramp-up component at time t; β represents the system's base load power; β represents the amplitude coefficient of the load's periodic fluctuations; α is the angular frequency parameter; φ is the initial phase angle; This represents the summation of the temperature influence terms, where k is the index of the temperature influence factor. The weighting coefficient for the k-th temperature factor is... Let k be the type of temperature-related variable at time t. The random disturbance term representing load ramp-up;

[0030] Wind power ramping component: Modeling is based on the physical relationship between the power curve of the wind turbine and the wind speed gradient;

[0031] In the formula, The wind power ramp-up component at time t represents the rate of change of wind power output with respect to time, and the unit is MW / h; The rated power of a wind turbine is its maximum output power at rated wind speed, expressed in MW. This indicates the wind speed at the current moment, in m / s. This represents the rate of change of wind speed at time t; the negative sign indicates that as the wind speed increases, the wind power output increases, corresponding to the positive ramping demand.

[0032] Solar ramping components: Modeling the impact of changes in light intensity on photovoltaic power output based on the rate of change of capacity factor;

[0033] In the formula, The photovoltaic ramp-up component at time t represents the rate of change of photovoltaic output with respect to time, expressed in MW / h. The rated installed capacity of the photovoltaic system is the maximum output power under standard test conditions, expressed in MW; CF represents the capacity factor, which is the dimensionless ratio of actual output to rated capacity, with a value range of [0, 1]. It represents the rate of change of capacity factor with respect to time, in units of 1 / h; the negative sign indicates that when the capacity factor increases, i.e., when the light intensity increases, the photovoltaic output increases, corresponding to the positive ramp-up demand.

[0034] Coupled components: Represents the interaction terms between the sources;

[0035] The state vector can be represented as:

[0036] ;

[0037] In the formula, This represents the system state vector at time t, which contains key information about the system's operation and is used as input for subsequent prediction models. Let be the load ramp-up component at time t; Let be the wind power ramping component at time t; Let be the photovoltaic ramp-up component at time t; The load power at time t is expressed in MW. The unit for wind power output at time t is MW; The photovoltaic output at time t is expressed in MW. This indicates that weather-related variables include multi-dimensional meteorological data such as temperature, humidity, air pressure, wind speed, wind direction, cloud cover, and irradiance. These represent time-related variables, including time characteristics such as hours, days of the week, months, and seasons, used to capture the periodic patterns of load.

[0038] Further, in step S2, constructing the corresponding input features includes:

[0039] For the load ramp component, we focus on extracting its time, day, week, month, and seasonal periodic characteristics;

[0040] For the wind power ramp component, we focus on extracting its randomness characteristics related to wind speed, wind direction and air pressure;

[0041] For the photovoltaic ramping component, we focus on extracting its intermittent characteristics related to solar irradiance, cloud cover, and temperature.

[0042] Further, in step S3, the processed component data are input into the corresponding differential LSTM branch networks, and the predicted feature vectors of each component are calculated in parallel, including:

[0043] The processed load ramp component data is input into the first LSTM branch network, which is a three-layer LSTM structure and uses the tanh activation function to extract the periodic temporal features of the load.

[0044] The processed wind power ramp component data is input into the second LSTM branch network, which is a two-layer LSTM structure and uses the ReLU activation function and layer normalization to capture the random fluctuation characteristics of wind power.

[0045] The processed photovoltaic ramping component data is input into the third LSTM branch network, which is a two-layer LSTM structure and uses the Swish activation function and gated linear unit mechanism to model the intermittent output characteristics of photovoltaics.

[0046] Each branch network performs computation in parallel and outputs its corresponding predicted feature vector.

[0047] Further, in step S4, the output of the final system total ramp demand prediction result includes:

[0048] ;

[0049] ;

[0050] in, This represents the output characteristics of each branch network. A context vector representing information such as renewable energy penetration rate, load factor, weather type, and season. This represents the dynamically calculated attention weights. This represents the predicted features of the final fusion.

[0051] Furthermore, in step S5, the multi-objective loss function includes:

[0052] ;

[0053] in, This indicates the overall forecast loss in ramp-up demand. This represents the weighted sum of the prediction losses for each component, where λ1, λ2, and λ3 represent the weighting coefficients. This represents a physical consistency constraint.

[0054] Compared with existing technologies, the advantages of this invention are as follows: Through fractional modeling, the system can perform specialized predictions for the periodicity of load, the randomness of wind power, and the intermittency of photovoltaic power, fully utilizing the inherent laws of each component. Regardless of the level of renewable energy penetration or changes in operating conditions, the system can achieve accurate predictions through a differentiated network architecture. Fractional modeling decomposes complex system ramp-up requirements into physically interpretable independent components, providing real-time analysis of the contribution of each source to the total ramp-up. Existing prediction systems require extensive debugging work in initial design and parameter configuration, and their adaptive capabilities are limited once deployed. This invention, by introducing a context-aware dynamic attention fusion mechanism, achieves dynamic operating condition awareness and adaptive optimization capabilities for the system. The system can automatically adjust the weight distribution of each component according to the real-time operating status, maintaining optimal prediction performance under different operating scenarios. Existing technologies often focus only on the single objective of prediction accuracy when optimizing prediction performance, while this invention achieves multi-objective collaborative optimization of prediction accuracy, physical consistency, and computational efficiency through a unified training framework. The fractional collaborative prediction algorithm of this invention is not only applicable to the current power system prediction scenario, but also has high engineering scalability, and can flexibly adapt to power grid systems of different sizes and diverse prediction needs. Compared with existing fixed-architecture prediction methods, the modular design algorithm in this invention can flexibly adjust the prediction strategy according to different system characteristics and business needs.

[0055] Furthermore, this invention employs a multi-head attention mechanism, where the context vector includes information on renewable energy penetration rate, system load rate, weather type, and season. This enables the system to dynamically perceive changes in grid operating conditions and adaptively adjust the fusion weights of each component. Consequently, it can maintain stable prediction accuracy and robustness under different renewable energy penetration rates and extreme weather conditions, thereby enhancing the system's environmental adaptability.

[0056] Furthermore, by designing proprietary network structures and activation functions for load, wind power, and photovoltaics respectively, this invention can accurately match their unique temporal characteristics (such as periodicity, randomness, and intermittency), thereby fully exploring the inherent variation patterns of each component and significantly improving the prediction accuracy of each component and the system as a whole.

[0057] Furthermore, the target loss function of the present invention includes the total prediction loss, the weighted sum of component prediction losses, and a physical consistency constraint term. By applying the physical consistency constraint during training, the sum of the prediction values ​​of each component is ensured to be physically and logically consistent with the overall prediction value, thus avoiding the problem of component predictions being disconnected from the total system value and enhancing the physical rationality and reliability of the prediction results.

[0058] Furthermore, based on physical mechanisms, this invention decomposes the total ramp demand of the system into three independent components: load, wind power, and photovoltaic power. It also provides specific mathematical decomposition formulas and physical modeling methods for each component, transforming the complex "black box" prediction problem into multiple sub-problems with clear physical meaning. This makes the prediction process transparent, quantitatively outputs the contribution of each energy source, and provides a direct and reliable scientific basis for power grid dispatching decisions.

[0059] Furthermore, this invention focuses on extracting periodic features for the load component, random features for the wind power component, and intermittent features for the photovoltaic component. By customizing feature engineering strategies for different components, it ensures that the feature vectors input to the corresponding branch networks are the most representative, thereby improving the model's learning efficiency and feature utilization from the source and laying a data foundation for high-precision prediction.

[0060] Furthermore, this invention inputs the data of each component into the corresponding differentiated LSTM branch network for parallel computation. By utilizing the parallel computing architecture, the three branch networks can be processed simultaneously, which greatly shortens the time required for model training and online prediction, and meets the second-level response requirements of real-time power grid dispatching.

[0061] Furthermore, this invention uses an attention mechanism formula for weighted fusion, which can dynamically quantify the importance of each component's prediction result based on the real-time context, achieving intelligent weighting rather than fixed weights. This allows the optimal fusion strategy to be automatically found in complex and ever-changing operating scenarios, further improving the accuracy of the final prediction result.

[0062] Furthermore, the multi-objective loss function of this invention uniformly optimizes the three objectives of overall system prediction accuracy, prediction accuracy of each component, and physical consistency. By using weighting coefficients, it achieves a balance among the multiple objectives, thereby training a prediction model with better overall performance and more in line with actual engineering needs. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the structure of the multi-source ramp-up collaborative prediction system of the pumped storage power system of the present invention;

[0064] Figure 2 This is a flowchart of the multi-source ramp-climbing collaborative prediction modeling method for pumped storage power systems according to the present invention. Detailed Implementation

[0065] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0066] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0067] Please see Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of the multi-source ramp-up collaborative prediction system of the pumped storage power system of the present invention.

[0068] The technical solution provided in this application includes the following steps:

[0069] The component decomposition module is used to decompose the total ramp demand of the power system into load ramp components, wind power ramp components, and photovoltaic ramp components based on the physical mechanism of the power system.

[0070] The differentiated parallel prediction module, connected to the component decomposition module, includes a first LSTM branch network for processing the load ramping component, a second LSTM branch network for processing the wind power ramping component, and a third LSTM branch network for processing the photovoltaic ramping component. The first LSTM branch network, the second LSTM branch network, and the third LSTM branch network adopt differentiated network architectures based on the time-series characteristics of their respective components.

[0071] The context-aware dynamic fusion module, connected to the differentiated parallel prediction module, is used to receive the output features of each LSTM branch network, and based on the context vector containing the system operating state, dynamically calculate the weight of each component through an attention mechanism, perform weighted fusion, and generate the final predicted value of the total system ramp demand.

[0072] This invention utilizes fractional modeling, enabling the system to perform specialized predictions for the periodicity of load, the randomness of wind power, and the intermittency of photovoltaic power, fully leveraging the inherent patterns of each component. Regardless of the level of renewable energy penetration or changes in operating conditions, the system can achieve accurate predictions through a differentiated network architecture. By fractional modeling, the complex system ramp-up requirements are decomposed into physically interpretable independent components, providing real-time analysis of the contribution of each source to the total ramp-up. Existing prediction systems require extensive debugging work in initial design and parameter configuration, and their adaptive capabilities are limited once deployed. This invention introduces a context-aware dynamic attention fusion mechanism to achieve dynamic operating condition awareness and adaptive optimization capabilities for the system. The system can automatically adjust the weight distribution of each component according to the real-time operating status, maintaining optimal prediction performance under different operating scenarios. Existing technologies often focus only on the single objective of prediction accuracy when optimizing prediction performance, while this invention achieves multi-objective collaborative optimization of prediction accuracy, physical consistency, and computational efficiency through a unified training framework. The fractional collaborative prediction algorithm of this invention is not only applicable to the current power system prediction scenario, but also has high engineering scalability, and can flexibly adapt to power grid systems of different sizes and diverse prediction needs. Compared with existing fixed-architecture prediction methods, the modular design algorithm in this invention can flexibly adjust the prediction strategy according to different system characteristics and business needs.

[0073] This invention employs a multi-head attention mechanism, with the context vector containing information on renewable energy penetration rate, system load rate, weather type, and season. This enables the system to dynamically perceive changes in grid operating conditions and adaptively adjust the fusion weights of each component. As a result, the system can maintain stable prediction accuracy and robustness under different renewable energy penetration rates and extreme weather conditions, thereby enhancing the system's environmental adaptability.

[0074] Specifically, the first LSTM branch network, the second LSTM branch network, and the third LSTM branch network adopt differentiated network architectures based on the temporal characteristics of their respective components.

[0075] The first LSTM branch network is a three-layer LSTM structure and uses the tanh activation function;

[0076] The second LSTM branch network is a two-layer LSTM structure, and uses the ReLU activation function and layer normalization;

[0077] The third LSTM branch network is a two-layer LSTM structure, and it adopts the Swish activation function and gated linear unit mechanism.

[0078] In this embodiment of the invention, three specialized LSTM branch networks are specifically designed as follows: Load branch network: 3-layer LSTM (64-48-32 units) + tanh activation + Dropout (0.2) specifically adapted to the daily, weekly, and seasonal variation patterns of loads; Wind power branch network: 2-layer LSTM (128 units) + ReLU activation + layer normalization + Dropout (0.3) specifically handled the high-frequency random fluctuations and intermittent characteristics of wind power; Photovoltaic branch network: 2-layer LSTM (96 units) + Swish activation + GLU mechanism + Dropout (0.25) specifically modeled the intermittent and asymmetric output characteristics of photovoltaics; The action vector can be represented as: ;

[0079] In the formula, This represents the set of output feature vectors of the three LSTM branch networks at time t. The output feature vector of the load branch network at time t is given. This vector is processed through a 3-layer LSTM structure and the tanh activation function is used to extract the daily periodicity, weekly periodicity, and seasonal variation patterns of the load. This is the output feature vector of the wind power branch network at time t. This vector, after passing through a two-layer LSTM structure and being processed with the ReLU activation function and layer normalization, captures the high-frequency random fluctuations and intermittent characteristics of wind power. Let be the output feature vector of the photovoltaic branch network at time t. This vector is processed by a 2-layer LSTM structure and the Swish activation function and gated linear unit mechanism to model the intermittent and asymmetric output features of the photovoltaic network.

[0080] This invention designs proprietary network structures and activation functions for load, wind power, and photovoltaic power respectively, which can accurately match their unique time-series characteristics (such as periodicity, randomness, and intermittency), thereby fully exploring the inherent variation patterns of each component and significantly improving the prediction accuracy of each component and the system as a whole.

[0081] Specifically, the context-aware dynamic fusion module adopts a multi-head attention mechanism, and the context vector includes one or more of the following: new energy penetration rate, system load rate, weather type, and seasonal information.

[0082] Specifically, the multi-source hill-climbing collaborative prediction system also includes a unified training and optimization framework, which uses a multi-objective loss function to train the system. The multi-objective loss function includes:

[0083] Total system ramp demand prediction loss;

[0084] The weighted sum of the predicted losses for each component;

[0085] The physical consistency constraint is used to ensure consistency between the sum of the predicted values ​​of each component and the predicted value of the total ramp demand of the system.

[0086] The target loss function of this invention includes the total prediction loss, the weighted sum of component prediction losses, and a physical consistency constraint term. By applying the physical consistency constraint during training, it ensures that the sum of the prediction values ​​of each component is consistent with the overall prediction value in terms of physical logic, avoiding the problem of component predictions being disconnected from the total system value, and enhancing the physical rationality and reliability of the prediction results.

[0087] Please see Figure 2 As shown, Figure 2 This is a flowchart of the multi-source ramp-climbing collaborative prediction modeling method for pumped storage power systems according to the present invention.

[0088] Specifically, a modeling method for a multi-source ramp-up collaborative prediction system applied to the pumped-storage power system includes:

[0089] Step S1: Collect multi-source operation data of the power system and decompose the total ramp demand of the system into load ramp component, wind power ramp component and photovoltaic ramp component based on physical mechanism;

[0090] Step S2: Perform specialized feature engineering on each component to construct the corresponding input features;

[0091] Step S3: Input the processed component data into the corresponding differential LSTM branch network and calculate the predicted feature vector of each component in parallel.

[0092] Step S4: Construct the system running context vector, dynamically calculate the fusion weight of the predicted features of each component using the attention mechanism, perform weighted fusion, and output the final prediction result of the total system ramp demand.

[0093] Step S5: Based on the prediction error feedback, update the network parameters using a multi-objective loss function.

[0094] Specifically, in step S1, the total system ramp demand is decomposed into load ramp component, wind power ramp component, and photovoltaic ramp component based on physical mechanisms, including:

[0095] The system's ramp-up requirement decomposition formula is as follows:

[0096] ;

[0097] The physical modeling methods for each component are as follows:

[0098] Load ramp component: Modeling is based on the periodic characteristics of the load and the influence of temperature;

[0099] In the formula, This represents the load ramp-up component at time t, i.e., the rate of change of load power with respect to time, expressed in MW / h. α represents the base load growth coefficient, reflecting the rate of long-term load trend growth. This represents the system's base load power, i.e., the average load level without considering periodic fluctuations and random disturbances, expressed in MW. β represents the amplitude coefficient of load periodic fluctuations, reflecting the intensity of intraday load fluctuations. The term is a sine function used to characterize the periodicity of the load, where α is the angular frequency parameter, reflecting the period length of the load, and φ is the initial phase angle, reflecting the time offset of the load peak. This represents the summation of the temperature influence terms, where k is the index of the temperature influence factor. The weighting coefficient for the k-th temperature factor. Let k be the type of temperature-related variable at time t. This term captures the impact of temperature changes on load ramp-up. The term representing the random disturbance of load ramping reflects the unpredictable load fluctuations following a random distribution with a mean of zero.

[0100] Wind power ramping component: Modeling is based on the physical relationship between the power curve of the wind turbine and the wind speed gradient;

[0101] In the formula, The wind power ramp component at time t, i.e., the rate of change of wind power output with respect to time, is expressed in MW / h. The rated power of a wind turbine is expressed as the maximum output power of the wind turbine at its rated wind speed, and the unit is MW. The wind speed at the current moment is expressed in m / s; The wind speed gradient represents the rate of change of wind speed at time t. This term directly determines the speed and direction of wind turbine ramping. The negative sign indicates that the wind power output increases when the wind speed increases, corresponding to the positive ramping demand. This formula is based on the physical relationship between the wind turbine power curve and the wind speed gradient.

[0102] Solar ramping components: Modeling the impact of changes in light intensity on photovoltaic power output based on the rate of change of capacity factor;

[0103] In the formula, The photovoltaic ramp-up component at time t, i.e., the rate of change of photovoltaic output with respect to time, is expressed in MW / h. The rated installed capacity of the photovoltaic system is the maximum output power under standard test conditions, and the unit is MW; CF represents the capacity factor, which is the ratio of actual output to rated capacity. It is dimensionless and its value ranges from [0, 1]. This parameter comprehensively reflects the combined influence of various factors such as solar irradiance, solar altitude angle, cloud cover, and temperature on photovoltaic output. The rate of change of the capacity factor with respect to time reflects the impact of changes in light intensity on photovoltaic output, with the unit being 1 / h; the negative sign indicates that when the capacity factor increases, i.e., when the light intensity increases, the photovoltaic output increases, corresponding to the positive ramp-up demand. This formula is based on modeling the impact of changes in light intensity on photovoltaic output by reflecting the rate of change of the capacity factor.

[0104] Coupled components: Represents the interaction terms between the sources;

[0105] The state vector can be represented as:

[0106] ;

[0107] In the formula, This represents the system state vector at time t, which contains key information about the system's operation and is used as input for subsequent prediction models. Let be the load ramp-up component at time t; Let be the wind power ramping component at time t; Let be the photovoltaic ramp-up component at time t; The load power at time t is expressed in MW. The unit for wind power output at time t is MW; The photovoltaic output at time t is expressed in MW. This indicates that weather-related variables include multi-dimensional meteorological data such as temperature, humidity, air pressure, wind speed, wind direction, cloud cover, and irradiance. These represent time-related variables, including time characteristics such as hours, days of the week, months, and seasons, used to capture the periodic patterns of load.

[0108] This invention decomposes the total ramp demand of the system into three independent components: load, wind power, and photovoltaic power, based on physical mechanisms. It also provides specific mathematical decomposition formulas and physical modeling methods for each component, transforming the complex "black box" prediction problem into multiple sub-problems with clear physical meaning. This makes the prediction process transparent and can quantitatively output the contribution of each energy source, providing a direct and reliable scientific basis for power grid dispatching decisions.

[0109] Specifically, in step S2, constructing the corresponding input features includes:

[0110] For the load ramp component, we focus on extracting its time, day, week, month, and seasonal periodic characteristics;

[0111] For the wind power ramp component, we focus on extracting its randomness characteristics related to wind speed, wind direction and air pressure;

[0112] For the photovoltaic ramping component, we focus on extracting its intermittent characteristics related to solar irradiance, cloud cover, and temperature.

[0113] This invention focuses on extracting periodic features for the load component, random features for the wind power component, and intermittent features for the photovoltaic component. By customizing feature engineering strategies for different components, it ensures that the feature vectors input to the corresponding branch networks are the most representative, thereby improving the model's learning efficiency and feature utilization from the source and laying a data foundation for high-precision prediction.

[0114] Specifically, in step S3, the processed component data are input into the corresponding differential LSTM branch networks, and the predicted feature vectors of each component are calculated in parallel, including:

[0115] The processed load ramp component data is input into the first LSTM branch network, which is a three-layer LSTM structure and uses the tanh activation function to extract the periodic temporal features of the load.

[0116] The processed wind power ramp component data is input into the second LSTM branch network, which is a two-layer LSTM structure and uses the ReLU activation function and layer normalization to capture the random fluctuation characteristics of wind power.

[0117] The processed photovoltaic ramping component data is input into the third LSTM branch network, which is a two-layer LSTM structure and uses the Swish activation function and gated linear unit mechanism to model the intermittent output characteristics of photovoltaics.

[0118] Each branch network performs computation in parallel and outputs its corresponding predicted feature vector.

[0119] This invention inputs the data of each component into the corresponding differentiated LSTM branch network for parallel computation. By utilizing the parallel computing architecture, the three branch networks can be processed simultaneously, which greatly shortens the time required for model training and online prediction, and meets the second-level response requirements of real-time power grid dispatch.

[0120] Specifically, in step S4, the output of the final system total ramp demand prediction result includes:

[0121] ;

[0122] ;

[0123] in, This represents the output characteristics of each branch network. A context vector representing information such as renewable energy penetration rate, load factor, weather type, and season. This represents the dynamically calculated attention weights. This represents the predicted features of the final fusion.

[0124] This invention uses an attention mechanism formula for weighted fusion, which can dynamically quantify the importance of each component's prediction result based on the real-time context, achieving intelligent weighting rather than fixed weights. This allows the system to automatically find the optimal fusion strategy in complex and ever-changing operating scenarios, further improving the accuracy of the final prediction result.

[0125] Specifically, in step S5, the multi-objective loss function includes:

[0126] ;

[0127] in, This indicates the overall forecast loss in ramp-up demand. This represents the weighted sum of the prediction losses for each component, where λ1, λ2, and λ3 represent the weighting coefficients. This represents a physical consistency constraint.

[0128] Specifically, in this embodiment, λ1, λ2, and λ3 represent weight coefficients used to balance the relative importance of different loss terms. Their determination method includes grid search based on the validation set, attempting different weight combinations on the validation set to select the parameter configuration that optimizes the overall performance index. Initial values ​​are set based on engineering experience, typically λ1 ∈ [0.3, 0.7] to ensure the accuracy of component prediction, λ2 ∈ [0.1, 0.5] to ensure physical consistency, and λ3 ∈ [0.01, 0.1] to prevent overfitting. A dynamic adjustment strategy is adopted: a smaller λ2 is used in the early stages of training to prioritize learning the features of each component, and λ2 is gradually increased in the later stages of training to strengthen physical constraints. regularization This represents a regularization term used to prevent overfitting of the model. L2 regularization is commonly used, and this term penalizes excessively large network weights to improve the model's generalization ability.

[0129] The multi-objective loss function of this invention uniformly optimizes three objectives: overall prediction accuracy, prediction accuracy of each component, and physical consistency. By using weighting coefficients, it achieves a balance among the multiple objectives, thereby training a prediction model with better overall performance and that better meets the actual needs of engineering.

[0130] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this application.

Claims

1. A multi-source ramp-up collaborative prediction system including a pumped-storage power system, characterized in that, include: The component decomposition module is used to decompose the total ramp demand of the power system into load ramp components, wind power ramp components, and photovoltaic ramp components based on the physical mechanism of the power system. The differentiated parallel prediction module, connected to the component decomposition module, includes a first LSTM branch network for processing the load ramping component, a second LSTM branch network for processing the wind power ramping component, and a third LSTM branch network for processing the photovoltaic ramping component. The first LSTM branch network, the second LSTM branch network, and the third LSTM branch network adopt differentiated network architectures based on the time-series characteristics of their respective components. The context-aware dynamic fusion module, connected to the differentiated parallel prediction module, is used to receive the output features of each LSTM branch network, and based on the context vector containing the system operating state, dynamically calculate the weight of each component through an attention mechanism, perform weighted fusion, and generate the final predicted value of the total system ramp demand.

2. The multi-source ramp-up collaborative prediction system including pumped storage power systems according to claim 1, characterized in that, The first, second, and third LSTM branch networks employ differentiated network architectures based on the temporal characteristics of their respective components. The first LSTM branch network is a three-layer LSTM structure and uses the tanh activation function; The second LSTM branch network is a two-layer LSTM structure, and uses the ReLU activation function and layer normalization; The third LSTM branch network is a two-layer LSTM structure, and it adopts the Swish activation function and gated linear unit mechanism.

3. The multi-source ramp-up collaborative prediction system including pumped storage power systems according to claim 1, characterized in that, The context-aware dynamic fusion module adopts a multi-head attention mechanism, and the context vector includes one or more of the following: new energy penetration rate, system load rate, weather type, and seasonal information.

4. The multi-source ramp-up collaborative prediction system including pumped storage power systems according to claim 1, characterized in that, The multi-source hill-climbing collaborative prediction system also includes a unified training and optimization framework, which trains the system using a multi-objective loss function, including: Total system ramp demand prediction loss; The weighted sum of the predicted losses for each component; The physical consistency constraint is used to ensure consistency between the sum of the predicted values ​​of each component and the predicted value of the total ramp demand of the system.

5. A multi-source ramp-climbing collaborative prediction modeling method for pumped-storage power systems, characterized in that, include: Step S1: Collect multi-source operation data of the power system and decompose the total ramp demand of the system into load ramp component, wind power ramp component and photovoltaic ramp component based on physical mechanism; Step S2: Perform specialized feature engineering on each component to construct the corresponding input features; Step S3: Input the processed component data into the corresponding differential LSTM branch network and calculate the predicted feature vector of each component in parallel. Step S4: Construct the system running context vector, dynamically calculate the fusion weight of the predicted features of each component using the attention mechanism, perform weighted fusion, and output the final prediction result of the total system ramp demand. Step S5: Based on the prediction error feedback, update the network parameters using a multi-objective loss function.

6. The multi-source ramp-climbing collaborative prediction modeling method for pumped-storage power systems according to claim 5, characterized in that, In step S1, the total system ramp demand is decomposed into load ramp component, wind power ramp component, and photovoltaic ramp component based on physical mechanisms, including: The system's ramp-up requirement decomposition formula is as follows: ； The physical modeling methods for each component are as follows: Load ramp component: Modeling is based on the periodic characteristics of the load and the influence of temperature; In the formula, This represents the load ramp-up component at time t; β represents the system's base load power; β represents the amplitude coefficient of the load's periodic fluctuations; α is the angular frequency parameter; φ is the initial phase angle; This represents the summation of the temperature influence terms, where k is the index of the temperature influence factor. The weighting coefficient for the k-th temperature factor is... Let k be the type of temperature-related variable at time t. The random disturbance term representing load ramp-up; Wind power ramping component: Modeling is based on the physical relationship between the power curve of the wind turbine and the wind speed gradient; In the formula, The wind power ramp-up component at time t represents the rate of change of wind power output with respect to time, and the unit is MW / h; The rated power of a wind turbine is its maximum output power at rated wind speed, expressed in MW. This indicates the wind speed at the current moment, in m / s. This represents the rate of change of wind speed at time t; the negative sign indicates that as the wind speed increases, the wind power output increases, corresponding to the positive ramping demand. Solar ramping components: Modeling the impact of changes in light intensity on photovoltaic power output based on the rate of change of capacity factor; In the formula, The photovoltaic ramp-up component at time t represents the rate of change of photovoltaic output with respect to time, expressed in MW / h. The rated installed capacity of the photovoltaic system is the maximum output power under standard test conditions, expressed in MW; CF represents the capacity factor, which is the dimensionless ratio of actual output to rated capacity, with a value range of [0, 1]. It represents the rate of change of capacity factor with respect to time, in units of 1 / h; the negative sign indicates that when the capacity factor increases, i.e., when the light intensity increases, the photovoltaic output increases, corresponding to the positive ramp-up demand. Coupled components: Represents the interaction terms between the sources; The state vector can be represented as: ； In the formula, This represents the system state vector at time t, which contains key information about the system's operation and is used as input for subsequent prediction models. Let be the load ramp-up component at time t; Let be the wind power ramping component at time t; Let be the photovoltaic ramp-up component at time t; The load power at time t is expressed in MW. The unit for wind power output at time t is MW; The photovoltaic output at time t is expressed in MW. This indicates that weather-related variables include multi-dimensional meteorological data such as temperature, humidity, air pressure, wind speed, wind direction, cloud cover, and irradiance. These represent time-related variables, including time characteristics such as hours, days of the week, months, and seasons, used to capture the periodic patterns of load.

7. The multi-source ramp-climbing collaborative prediction modeling method for pumped-storage power systems according to claim 5, characterized in that, In step S2, constructing the corresponding input features includes: For the load ramp component, we focus on extracting its time, day, week, month, and seasonal periodic characteristics; For the wind power ramp component, we focus on extracting its randomness characteristics related to wind speed, wind direction and air pressure; For the photovoltaic ramping component, we focus on extracting its intermittent characteristics related to solar irradiance, cloud cover, and temperature.

8. The multi-source ramp-climbing collaborative prediction modeling method for pumped storage power systems according to claim 5, characterized in that, In step S3, the processed component data are input into the corresponding differential LSTM branch networks, and the predicted feature vectors of each component are calculated in parallel, including: The processed load ramp component data is input into the first LSTM branch network, which is a three-layer LSTM structure and uses the tanh activation function to extract the periodic temporal features of the load. The processed wind power ramp component data is input into the second LSTM branch network, which is a two-layer LSTM structure and uses the ReLU activation function and layer normalization to capture the random fluctuation characteristics of wind power. The processed photovoltaic ramping component data is input into the third LSTM branch network, which is a two-layer LSTM structure and uses the Swish activation function and gated linear unit mechanism to model the intermittent output characteristics of photovoltaics. Each branch network performs computation in parallel and outputs its corresponding predicted feature vector.

9. The multi-source ramp-climbing collaborative prediction modeling method for pumped storage power systems according to claim 5, characterized in that, In step S4, the final predicted total system ramp demand is output, including: ； ； in, This represents the output characteristics of each branch network. A context vector representing information on renewable energy penetration rate, load factor, weather type, and season. This represents the dynamically calculated attention weights. This represents the predicted features of the final fusion.

10. The multi-source ramp-climbing collaborative prediction modeling method for pumped-storage power systems according to claim 5, characterized in that, In step S5, the multi-objective loss function includes: ； in, This indicates the overall forecast loss in ramp-up demand. This represents the weighted sum of the prediction losses for each component, where λ1, λ2, and λ3 represent the weighting coefficients. This represents a physical consistency constraint.

Images