A new energy combination prediction and optimization method in a new power system based on a multi-neural network

Abstract

The application discloses a new energy combination prediction and optimization method in a new power system based on a multi-neural network, which comprises the following steps: S1, a CNN-LSTM-GWO combination prediction model is built, the model is compared with a single LSTM neural network and a CNN-LSTM combination prediction model, training is carried out by inputting the same training set, and the prediction effects of test sets are compared; and S2, the prediction results of different models are used in subsequent optimization scheduling engineering of the new power system, a weight updating intraday rolling optimization method is used in the optimization scheduling process of the new power system, and the cost of the new power system is compared between the method and a traditional rolling optimization method. The application explores the latest progress by integrating two particularly effective neural networks for predicting wind energy and solar energy. The combination improves the accuracy of model prediction. In addition, the training process is accelerated by combining the grey wolf optimization algorithm, and the practicability of the model for ultra-short-term prediction and optimization scheduling tasks is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of renewable energy mass integration into new power systems, and particularly relates to a new energy combination prediction and optimization method in a new power system based on a multi-neural network. BACKGROUND

[0002] With the mass integration of renewable energy into new power systems, its inherent randomness and volatility pose major challenges to optimization and scheduling. Therefore, accurate prediction of solar and wind power generation, as well as effective management of prediction data in the subsequent optimization process of new power systems, becomes increasingly critical. In the field of wind and solar power prediction, the progress of big data and artificial intelligence technology has led to the emergence of numerous modern methods, mainly including machine learning and deep learning techniques. These methods have strong data processing capabilities and the ability to model nonlinear relationships, and have gradually become the mainstream prediction method. Machine learning-based load prediction methods include linear regression, support vector machines, artificial neural networks (ANN). With the increasing complexity of data, some deep learning algorithm models have emerged as promising alternatives. Recurrent neural networks (RNN) are a deep learning model in which the output of RNN depends on the previous input, taking advantage of the time correlation in the sequence. In theory, RNN can model sequences of arbitrary length. Researchers have demonstrated significant performance improvements compared to other modeling methods by training a large-scale recurrent neural language model using nearly 1 billion words as training data. However, its accuracy is still not ideal in the face of more variable and complex prediction scenarios.

[0003] First of all, for wind and light prediction models, the main shortcomings of current machine learning methods are the inability to learn temporal features from input data, limited flexibility in identifying complex patterns, difficulty in capturing complex variable relationships, and susceptibility to overfitting problems. For neural networks, a single neural network is usually unable to perform comprehensive feature learning and information processing, and the model's generalization ability is limited when dealing with complex tasks and data. Combination prediction models developed on the basis of a single neural network can better solve the above problems. Combination prediction models mainly achieve this goal by combining individual models with different characteristics, thereby taking advantage of each model and mitigating the limitations of any single model. The present application uses a combination model of convolutional neural networks (CNN) and LSTMs to predict wind and light output, and the results show that the combination prediction model has the best prediction performance.

[0004] Secondly, for the fine scheduling of the day, the time of each scheduling window is very urgent, so the training process of the neural network used is not too long, otherwise it cannot meet the fine scheduling requirements of the day, and generally speaking, the training time of the neural network is from tens of minutes to several hours with the increase of data and the complexity of the scene. Based on this, it is urgent to optimize the training process of the neural network to speed up the training speed and training accuracy, and the present application starts from the hyperparameter setting of the neural network, and uses the grey wolf optimization algorithm (GWO, Grey wolf optimizer) to quickly lock the best training hyperparameter, and improves the training speed of the model.

[0005] Finally, as the prediction time range expands, any form of prediction model will show a larger and larger prediction error. In order to solve this problem, the present application proposes a new power system scheduling rolling optimization method combined with day-ahead and day-ahead weight updating. Compared with the current mainstream day-ahead scheduling method and the traditional rolling optimization method, the method improves the accuracy of the prediction data and reduces the overall scheduling cost. SUMMARY

[0006] The present application aims to provide a new type of new energy combination prediction and optimization method in a new power system based on a multi-neural network.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0008] The present application provides a new type of new energy combination prediction and optimization method in a new power system based on a multi-neural network, which comprises the following steps:

[0009] S1, build a CNN-LSTM-GWO combined prediction model, compare it with a single LSTM neural network and a CNN-LSTM combined prediction model, train by inputting the same training set, and compare the prediction effect of the test set;

[0010] S2, use the prediction results of different models in the subsequent optimization scheduling engineering of the new type of power system, use the day-ahead rolling optimization method of weight updating in the optimization scheduling process of the new type of power system, and compare the cost of the new type of power system of the method with that of the traditional rolling optimization method.

[0011] Further, in the step S2, it comprises:

[0012] A day-ahead scheduling optimization model is established: the purpose of day-ahead optimization scheduling is to minimize the operation cost of the new power system, which involves renewable energy power prediction based on the previous day's prediction, considering the peak and off-peak price difference, various traditional power generation costs, and energy storage charging and discharging costs, optimizing the discharge and charge behavior of the energy storage system and the power generation of each generator, and formulating the corresponding buying and selling strategy according to the price fluctuation;

[0013] A day-ahead scheduling optimization model is established: the purpose of day-ahead optimization scheduling is to minimize the operation cost of the new power system, which involves renewable energy power prediction based on the previous day's prediction, considering the peak and off-peak price difference, various traditional power generation costs, and energy storage charging and discharging costs, optimizing the discharge and charge behavior of the energy storage system and the power generation of each generator, and formulating the corresponding buying and selling strategy according to the price fluctuation;

[0014] Further, the day-ahead scheduling optimization model includes a day-ahead scheduling optimization objective function and system constraints, and the objective of the day-ahead scheduling optimization objective function is to reduce the operation cost of the system within the entire scheduling period.

[0015] Further, the day-ahead scheduling optimization objective function is as follows:

[0016]

[0017] In the formula, is the total economic cost of power purchase and sale, is the total operation cost of each generator, is the dispatching cost of the energy storage unit;

[0018]

[0019] In the formula, is the market electricity price in the time period, P t m is the new power system purchase and sale quantity in the t period;

[0020]

[0021] E(u n,t )=u n,t (1-u n,t-1 )S n (8)

[0022] In the formula, N G is the number of schedulable units, P g,n,t is the output of unit n in the t period, a n ,b n ,c n are the fuel cost coefficients of unit n, S n is the start-up cost of unit n, and u n,tdenotes the operation state of unit n at time period t, with value 1 indicating that the unit is running and value 0 indicating that the unit is off,

[0023]

[0024] where, is the unit cost of charging and discharging the energy storage battery; is the energy discharged and charged by the battery; N s is the number of energy storage batteries.

[0025] Further, the system constraints are specified as follows:

[0026]

[0027] where, is the discharging power of energy storage unit s at time period t, is the charging power of energy storage unit s at time period t, P F,f,t is the power of distributed energy source f at time period t, N f is the number of distributed energy sources. is the load power at time period t;

[0028]

[0029] where, and are the ramp-down and ramp-up rates of the unit, respectively;

[0030]

[0031] where, E se,s,t denotes the energy stored in storage unit s at time period t; the charging and discharging rates of the storage unit are denoted as and The unit also has specified maximum and minimum power constraints, denoted as and The storage unit s is limited by predefined upper and lower energy limits, denoted as and

[0032] 0≤P F,f,t ≤P max F,f,t (17)

[0033] where, P max F,f,t is the output limit of distributed energy source s.

[0034] Further, the objective function of the intra-day rolling optimization model is as follows:

[0035]

[0036] w t = 4*F t / FA t (22)

[0037] In the formula, r t represents the scheduling error between the day-ahead scheduling and the intra-day scheduling of the same unit at the same time, to enhance the global optimality of the intra-day fine scheduling, respectively represent the output of the energy storage unit, the output of the dispatchable unit, and the day-ahead scheduling of buying and selling electricity, respectively represent the output of the energy storage unit, the output of the dispatchable unit, the intra-day scheduling of buying and selling electricity, and u1, u2, u3 represent error coefficients, respectively represent the predicted output and the actual output of the renewable energy unit at t-4, and formula (21) and formula (22) define the updated weight value.

[0038] Compared with the prior art, the beneficial effects of the present application are:

[0039] In the context of the prediction model, the present application explores the recent progress by integrating two particularly effective neural networks for predicting wind energy and solar energy. This combination improves the accuracy of model prediction. In addition, the combination of the grey wolf optimization algorithm speeds up the training process and enhances the practicality of the model for ultra-short-term prediction and optimal scheduling tasks.

[0040] After improving the accuracy of the prediction itself, this paper proposes a weight updating method based on day-ahead collaborative rolling scheduling, which further solves the accuracy problem in subsequent optimal scheduling. This method not only retains the global optimality of traditional day-ahead scheduling, but also improves the accuracy of prediction data at each rolling optimization time step, thereby reducing the scheduling cost. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a diagram for rolling window optimization;

[0042] Figure 2 is a structural diagram of the convolutional neural network;

[0043] Figure 3 is a diagram of the constituent unit of LSTM. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0045] The application provides a new energy combination prediction and optimization method in a new power system based on a multi-neural network, comprising the following steps:

[0046] S1, a CNN-LSTM-GWO combination prediction model is built, which is compared with a single LSTM neural network and a CNN-LSTM combination prediction model, the same training set is input for training, and the prediction effects of the test sets are compared;

[0047] S2, the prediction results of different models are used in subsequent optimization scheduling engineering of the new power system, the weight updating day-ahead rolling optimization method is used in the optimization scheduling process of the new power system, and the cost of the new power system is compared between the method and the traditional rolling optimization method.

[0048] In step S2, the following steps are included:

[0049] A day-ahead scheduling optimization model is established: the purpose of day-ahead optimization scheduling is to minimize the operation cost of the new power system, which involves renewable energy power prediction based on the previous day prediction, considering the peak and off-peak price difference, various traditional power generation costs and energy storage charging and discharging costs, optimization includes scheduling the discharging and charging behavior of the energy storage system and the power generation capacity of each generator, and formulating the corresponding buying and selling strategy according to the price fluctuation;

[0050] A day-ahead rolling optimization model is established: rolling optimization is realized within a day, a prediction model is used to perform rolling prediction with a resolution of 15 minutes across 96 nodes, each rolling optimization prediction window spans 4 time steps, the first step of the optimal prediction sequence is executed, and then the rolling optimization step is repeated in turn.

[0051] The day-ahead scheduling optimization model includes a day-ahead scheduling optimization objective function and system constraints, and the objective of the day-ahead scheduling optimization objective function is to reduce the operation cost of the system within the entire scheduling period.

[0052] Specifically as follows:

[0053] Day-ahead optimization scheduling modeling

[0054] The purpose of day-ahead optimization scheduling is to minimize the operation cost of the new power system. This involves renewable energy power prediction based on the previous day prediction, considering the peak and off-peak price difference, various traditional power generation costs and energy storage charging and discharging costs. Optimization includes scheduling the discharging and charging behavior of the energy storage system and the power generation capacity of each generator, and formulating the corresponding buying and selling strategy according to the price fluctuation. The goal of this method is to minimize the operation cost of the system.

[0055] a) Objective function

[0056] The objective of the objective function is to reduce the operation cost of the system in the whole scheduling period.

[0057]

[0058] wherein is the total economic cost of electricity purchase and sale, is the total operation cost of each generator unit, is the scheduling cost of the energy storage unit.

[0059]

[0060] wherein, is the market electricity price in the time period, is the electricity purchase and sale amount of the new power system participating in the market in the t period.

[0061]

[0062] wherein N G is the number of schedulable units, P g,n,t is the output of unit n in the t period, a n ,b n ,c n is the fuel cost coefficient of unit n, S n is the start-up cost of unit n, u n,t represents the operation state of unit n in the t period, and the value of 1 indicates that the unit is running, and the value of 0 indicates that the unit is shut down.

[0063]

[0064] wherein, is the unit cost of charging and discharging of the energy storage battery. is the energy of the battery discharge and charge. N s is the number of energy storage batteries.

[0065] b) System constraints

[0066]

[0067] wherein, is the discharge power of the energy storage unit s in the period t, is the charging power of the energy storage unit s in the period t, P F,f,t is the power of the distributed energy f in the period t, N f is the number of distributed energy. is the load power of the t period.

[0068]

[0069] wherein, and are the unit's downward and upward ramp rates, respectively.

[0070]

[0071]

[0072] where E se,s,t denotes the stored energy of storage unit s in time period t. The charging and discharging rates of the storage unit are denoted as and The unit also has specified maximum and minimum power constraints, denoted as and In addition, the storage unit s is subject to predefined upper and lower energy limits, denoted as and

[0073] 0 < P F,f,t < P max F,f,t (17)

[0074] where P max F,f,t is the output limit of distributed energy source s.

[0075] C. Intra-day rolling optimization

[0076] The rolling optimization is implemented within a day, using a prediction model for rolling prediction across 96 nodes. Each rolling optimization prediction window spans 4 time steps, performing the first step of the optimal prediction sequence, followed by a looped rolling optimization step in turn.

[0077] The results of the advanced scheduling show better global optimality, but are associated with large errors in solar and wind output data. In contrast, the intra-day scheduling benefits from higher accuracy of solar and wind data, but can only achieve local optimal solutions in general. Therefore, in the objective function of the intra-day scheduling, the day-ahead error and the intra-day scheduling error are taken as penalty terms to enhance the global optimality of the intra-day scheduling results.

[0078] In addition, it is necessary to predict the error patterns of solar and wind data according to the neural network model, and to define the weights of different time periods within the rolling time domain. Generally, the prediction accuracy of the neural network will decrease as the time scale increases.

[0079] Taking the above two factors into consideration, the intra-day scheduling optimization model is established. The main difference from the day-ahead scheduling lies in the optimization adjustment of the objective function:

[0080]

[0081] wt = 4*F t / FA t (22)

[0082] It can be seen that in order to more accurately describe the objective function, two new parameters are defined here, r t and w t . Among them, r t represents the scheduling error between the day-ahead scheduling and the intra-day scheduling of the same unit at the same time, in order to enhance the global optimality of the intra-day fine scheduling, respectively represent the output of the energy storage unit, the output of the dispatchable unit, and the day-ahead scheduling of buying and selling electricity strategy, respectively represent the output of the energy storage unit, the output of the dispatchable unit, and the intra-day scheduling of buying and selling electricity strategy, and u1, u2, u3 represent error coefficients, respectively represent the predicted output and actual output of the renewable energy unit at t-4 time. In addition, formula (21) and formula (22) define the updated weight value.

[0083] In addition, the model principles of CNN and LSTM.

[0084] 1. CNN

[0085] CNN is based on the principle of neural perceptron, inspired by the animal visual system, and represents a special form of deep feedforward neural network. They are widely used in image processing and natural language processing and other fields. As shown in Figure 2 , the design of CNN is composed of convolutional layers, pooling layers and fully connected layers. The network first uses convolutional layers to extract features from input data, then enhances these features through pooling layers, and finally integrates the results through fully connected layers.

[0086] 2. LSTM

[0087] In order to study and predict the mapping relationship between wind and solar output and daily time series, consider using LSTM network from deep neural network, whose network state update process is as shown in Figure 3 .

[0088] In the above structure, the forget gate decides which information to discard. The forget gate reads the user's response from the previous time step h t-1 and the current input information x t , and outputs a number between 0 and 1. The update formula of the forget gate is as follows:

[0089] f t = σ(w f × [h t-1 , x t ]) + b f (1)

[0090] where b f is a bias associated with the forget gate, σ denotes a sigmoid function, w f denotes a weight of the forget gate, x t denotes an input at the current time step, h t-1, denotes an output at the previous time step.

[0091] The input gate decides which values will be used to update the state of the user response state keeping unit. It first processes the input vector using a sigmoid activation function while generating a candidate value vector using a tanh function. The update formula for the input gate is as follows:

[0092] i t = σ(w i x t-1 h t + b i ) (2)

[0093]

[0094] where w i and b i denote the weights and biases of the input gate, while w c and b c denote the weights and biases of the candidate value vector, denoted as

[0095] The user response state keeping unit maintains the state of the previous time step. After processing the current input data, it is necessary to update the state information in the cell. The data of the forget gate is used to select what to discard, while the data of the input gate is used to select what to update. The update formula for the user response state keeping unit is as follows:

[0096]

[0097] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and that the present application can be implemented in other particular forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, the scope of the present application being defined by the appended claims rather than the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference signs in the claims should not be construed as limiting the claims to the figures in which the reference signs are used.

[0098] Furthermore, it should be understood that although the specification is described in terms of embodiments, not every embodiment includes every feature or implementation described herein. The specification can include implicit combinations of explicitly mentioned features and / or implicit combinations of implicitly mentioned features. Such combinations are also expressly included within the scope of the specification and an embodiment.

Claims

1. A new method for forecasting and optimizing renewable energy combinations in power systems based on multiple neural networks, characterized by: The following steps are involved: S1. Build a CNN-LSTM-GWO combined prediction model and compare it with a single LSTM neural network and a CNN-LSTM combined prediction model. Train them by inputting the same training set and compare the prediction results on the test set. S2. Use the prediction results of different models in the subsequent optimization and dispatching projects of new power systems. Apply the weight-updated intraday rolling optimization method to the optimization and dispatching process of new power systems, and compare the cost of the new power systems with that of the traditional rolling optimization method. The step S2 includes: establishing a day-ahead scheduling optimization model and establishing an intraday rolling optimization model; The day-ahead scheduling optimization model includes a day-ahead scheduling optimization objective function and system constraints. The day-ahead scheduling optimization objective function aims to reduce the system operating cost within the entire scheduling cycle. The day-ahead scheduling optimization objective function is as follows: In the formula is the overall economic cost of purchasing and selling electricity in time period t, is the total operating cost of each generator set during time period t, is the dispatching cost of the energy storage unit in time period t; Where, is the market electricity price in time period t, P t m The amount of electricity purchased and sold by the new power system participating in the market during period t; E(in n,t )=in n,t (1-in n,t-1 )WITH n (8) Where N G is the number of dispatchable units, P g,n,t is the output of unit n during period t, a n ,b n ,c n is the fuel cost coefficient of unit n, S n is the startup cost of unit n, u n,t Indicates the operating status of unit n during period t. A value of 1 indicates that the unit is running, and a value of 0 indicates that the unit is shut down. Where, P is the unit cost of charging and discharging the energy storage battery in time period t; se,s,t is the energy discharged and charged by the sth battery in time period t; N s is the total number of energy storage batteries.

2. The method for forecasting and optimizing new energy combinations in a novel power system based on multiple neural networks according to claim 1 is characterized in that: The system constraints are as follows: Where, is the discharge power of energy storage unit s in time period t, is the charging power of energy storage unit s in time period t, P F,f,t is the power of distributed energy f in time period t, N f is the number of distributed energy resources; P t L is the load power in the tth period; Where, and are the downward and upward climbing rates of the unit respectively; Where, E se,s,t represents the energy storage of storage unit s in time period t; the charging rate and discharging rate of storage unit are expressed as and The unit also has specified maximum and minimum power constraints, denoted as and The storage unit s is subject to predefined upper and lower energy limits, expressed as and 0≤P F,f,t ≤P max F,f,t (17) Where, P max F,f,t is the output limit of the fth distributed energy in time period t.

3. The method for forecasting and optimizing new energy combinations in a novel power system based on multiple neural networks according to claim 1 is characterized in that: The objective function of the intraday rolling optimization model is as follows: w t =4*F t / FA t (22) Where r t represents the scheduling error between the day-ahead scheduling and the intraday scheduling of the same unit at the same time, in order to enhance the global optimality of the intraday fine scheduling, They represent the output of the energy storage unit, the output of the dispatchable unit, and the day-ahead dispatch power purchase and sales strategy. They represent the output of the energy storage unit, the output of the dispatchable unit, and the intraday dispatch power purchase and sales strategy, u1, u2, and u3 represent the error coefficients. represent the predicted output and actual output of the renewable energy unit at time t-4, respectively. Equations (21) and (22) define the updated weight values.

Images