Load prediction method and system combining maximum information coefficient method and deep learning

Abstract

The invention relates to the technical field of load prediction, in particular to a load prediction method and system combining a maximum information coefficient method and deep learning. The method comprises the following steps: screening main factors influencing the load of the hydraulic turbine set to obtain the maximum information coefficient of each influence factor; taking the preprocessed data as input data of a first model, and determining basic hyper-parameters of the first model according to the input data; optimizing the first model by using an intelligent optimization algorithm to perform feature extraction so as to improve the feature capturing capability; and predicting the load of the hydraulic turbine set, and outputting a load prediction result.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a load forecasting method and system that combines the maximum information coefficient method and deep learning. Background Technology

[0002] With the large-scale grid connection of new energy sources, especially solar and wind power, the load characteristics of the power system have indeed changed significantly. Traditional energy sources such as thermal power and hydropower have relatively stable output characteristics, while solar and wind power are directly affected by weather conditions (such as sunlight intensity and wind speed), exhibiting greater volatility and uncertainty. In particular, with the widespread application of distributed photovoltaic systems, load curve changes have become more complex, increasing the difficulty of power system load forecasting.

[0003] Distributed photovoltaic (PV) systems are installed at the user end, typically after the electricity meter, making direct monitoring of their output difficult. This intermittent and fluctuating PV output, coupled with the power system load, affects the grid's load characteristics, increases the uncertainty of load changes, and thus impacts the accuracy of load forecasting. To address this challenge, the construction of smart grids provides richer and more real-time means of acquiring load data. The application of smart meters and IoT technologies makes load data collection more accurate and frequent, reaching frequencies of per minute or even per second. Simultaneously, precise meteorological data and equipment status data obtained through sensors and remote sensing technologies also provide more comprehensive information for load forecasting.

[0004] These technological advancements have undoubtedly improved the accuracy of load forecasting, but they have also brought about an expansion in data volume. How to effectively process and analyze this massive amount of data and extract valuable information from it is a crucial issue facing current power system load forecasting. Therefore, the development of smart grids and data processing technologies not only enhances load forecasting capabilities but also makes the operation of power systems more intelligent and efficient. Summary of the Invention

[0005] The purpose of this invention is to provide a load forecasting method that combines the maximum information coefficient method and deep learning, which can achieve accurate forecasting of future loads, realize the rational allocation of power output of hydropower units and other power sources (such as thermal power, wind power, etc.), reduce power generation costs, and improve economic efficiency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a load prediction method combining the maximum information coefficient method and deep learning, comprising:

[0008] The main factors affecting the load of the turbine unit were screened, and the maximum information coefficient of each influencing factor was obtained.

[0009] The preprocessed data is used as the input data for the first model, and the basic hyperparameters of the first model are determined based on the input data.

[0010] The first model is optimized using intelligent optimization algorithms to improve feature extraction capabilities;

[0011] The load of the turbine unit is predicted, and the load prediction results are output.

[0012] As a preferred embodiment of the load forecasting method combining the maximum information coefficient method and deep learning described in this invention, the step of screening the main factors affecting the turbine unit load and obtaining the maximum information coefficient of each influencing factor includes:

[0013] The mutual information algorithm is used to calculate the correlation of factors affecting energy consumption;

[0014] The maximum mutual information values ​​on different networks are preprocessed to obtain the preprocessed feature matrix.

[0015] The highest preprocessed value obtained from the feature matrix is ​​the maximum information coefficient.

[0016] As a preferred embodiment of the load prediction method combining the maximum information coefficient method and deep learning described in this invention, wherein: the step of optimizing the first model using an intelligent optimization algorithm to improve feature extraction capability includes:

[0017] The first model hyperparameter combination is generated by randomly generating multiple sets of tree depths and minimum sample weights for leaf nodes;

[0018] Use cross-validation to evaluate the performance of each hyperparameter combination;

[0019] The optimal solution within the range is found using intelligent optimization algorithms;

[0020] Stop when the maximum number of iterations is reached or the fitness no longer improves;

[0021] The first model was trained with optimal hyperparameters and its performance was evaluated.

[0022] As a preferred embodiment of the load forecasting method combining the maximum information coefficient method and deep learning described in this invention, the step of forecasting the turbine unit load and outputting the load forecasting result includes:

[0023] The optimized parameters are assigned to the first model, and the combined model is used to predict the turbine unit load.

[0024] The model is tested using test set data, the load forecast results are output, and the evaluation index is calculated to assess the model performance.

[0025] As a preferred embodiment of the load prediction method combining the maximum information coefficient method and deep learning described in this invention, the intelligent optimization algorithm incorporates a cyclic population reduction technique. During the optimization process, this technique accelerates convergence by extracting a portion of the population's population cores (CPs) and then reintroducing individuals into the population. This can be expressed as:

[0026]

[0027] Where T is the variable that determines the number of iterations; T max N is the maximum number of loops; min The minimum number of individuals in the population; t represents the solution value of the function in the current iteration; N is the population size.

[0028] As a preferred embodiment of the load forecasting method combining the maximum information coefficient method and deep learning described in this invention, wherein: the mutual information algorithm is used to calculate the correlation of factors affecting energy consumption, expressed as:

[0029]

[0030] Wherein, given a dataset D and grid regions G, p(a,b) is the joint probability density of variables a and b, p(a) and p(b) are the marginal probability densities of variables a and b calculated by the histogram estimation method, respectively, and MI(D|G) is the mutual information of grid regions G on dataset D. For a fixed given dataset D, the normalized MI feature matrix is ​​obtained by normalizing the maximum mutual information values ​​on different grids G as follows:

[0031]

[0032] The maximum information coefficient is the highest normalized MI value obtained from the feature matrix, expressed as:

[0033]

[0034] By calculating the maximum mutual information coefficient, the energy consumption correlation factors of hydropower units with coefficients greater than the threshold are selected as the influencing factor combination sample set as the model input.

[0035] The historical data is normalized and represented as follows:

[0036]

[0037] Where e represents the historical data before normalization; E represents the historical data after normalization; e max e represents the maximum value in the historical data before normalization; min This represents the minimum value in the historical data before normalization.

[0038] As a preferred embodiment of the load forecasting method combining the maximum information coefficient method and deep learning described in this invention, the calculated evaluation index for assessing model performance is expressed as:

[0039]

[0040]

[0041] Among them, y t and y t ' and ' represent the actual and predicted load power values ​​at time t, respectively; This represents the average value of the actual load power; N is the number of samples used for testing; MSE represents the mean square error; RMSE represents the root mean square error; MAPE represents the mean absolute percentage error; R 2 It is expressed as the coefficient of determination.

[0042] Secondly, the present invention provides a load forecasting system combining the maximum information coefficient method and deep learning, comprising:

[0043] The module includes a screening module, a determination module, an optimization module, and a prediction module.

[0044] Thirdly, the present invention provides an electronic device, comprising:

[0045] Memory and processor;

[0046] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of a load prediction method combining the maximum information coefficient method and deep learning.

[0047] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the load prediction method combining the maximum information coefficient method and deep learning.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. In response to the increasing number of influencing factors that make prediction difficult, this invention reduces the number of influencing factors and further improves the accuracy of the model by extracting features and inputting them into the model.

[0050] 2. Utilize intelligent optimization algorithms to optimize the feature extraction model and improve the model's operating efficiency.

[0051] 3. By leveraging the advantages of combined models, we can further improve the fit between predicted and actual values. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating the load prediction method combining the maximum information coefficient method and deep learning, as described in one embodiment of the present invention.

[0054] Figure 2 This diagram illustrates the fitness and convergence comparison of different optimization algorithms for a load prediction method combining the maximum information coefficient method and deep learning, as described in an embodiment of the present invention.

[0055] Figure 3 This diagram illustrates the average standard deviation of various optimization algorithms in a load forecasting method combining the maximum information coefficient method and deep learning, as described in an embodiment of the present invention.

[0056] Figure 4 This diagram illustrates a comparison of different models of the load prediction method combining the maximum information coefficient method and deep learning, as described in one embodiment of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0058] Example 1, an embodiment of the present invention, provides a load prediction method combining the maximum information coefficient method and deep learning, including:

[0059] S1: Perform noise reduction processing on the historical operating data of the turbine units and divide the dataset into training and testing sets.

[0060] S2: Screen the main factors affecting the turbine unit load and obtain the maximum information coefficient of each factor. Use the maximum information coefficient method (MIC) to screen the main factors affecting the turbine unit load, obtain the maximum information coefficient value of each factor, and normalize the training set and test set data.

[0061] S3: Using the preprocessed data as input data for the first model, determine the basic hyperparameters of the first model based on the input data. Then, using the normalized data as input data for the XGBoost model, perform further feature filtering and determine the basic hyperparameters of the XGBoost model based on the input data.

[0062] S4: Utilize intelligent optimization algorithms to optimize the first model for feature extraction, improving feature capture capabilities. Addressing the issue of poor prediction results due to difficulties in hyperparameter tuning during feature selection, the CPO (Crowning Pig Optimization) algorithm is used to optimize the XGBoost model for feature extraction, enhancing feature capture capabilities, reducing manual intervention, and thus improving prediction accuracy.

[0063] S5: Predict the turbine unit load and output the load prediction results. The depth of the tree and the minimum sample weight of the leaf nodes in the XGBoost model are optimized using CPO. The optimized parameters are assigned to the model, and a combined model of temporal convolutional network and bidirectional long short-term memory network (TCN-BiLSTM) is used to predict the turbine unit load. The model is tested using test set data, and the load prediction results are output. Multiple evaluation indicators are calculated to assess the model performance.

[0064] It should be noted that, facing the increasing number of influencing factors leading to higher prediction difficulty, this invention reduces the number of influencing factors and further improves the model's accuracy by extracting features and inputting them into the model. Intelligent optimization algorithms are used to optimize the feature extraction model, improving its operating efficiency. Furthermore, by leveraging the advantages of combined models, the degree of fit between predicted and true values ​​is further enhanced.

[0065] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, a load prediction method combining the maximum information coefficient method and deep learning is provided, such as... Figure 1 As shown.

[0066] In this embodiment of the application, step S1 involves denoising the historical operating data of the turbine unit and dividing the dataset into a training set and a test set. The specific implementation process is as follows:

[0067] The data denoising method used is the Isolation Forest method, and the formula is as follows:

[0068] E(h(x))→c(n),s→0.5 (1)

[0069]

[0070] E(h(x))→n-1,s→0 (3)

[0071]

[0072] H(*)=ln(*)+ξ (5)

[0073] Where c(n) is a dataset containing n samples, representing the average path length of the tree, used to standardize the path length of record x; H(*) is the harmonic number, and ξ is Euler's constant, with an approximate value of 0.577.

[0074] In this embodiment of the application, step S2, which involves screening the main factors affecting the turbine unit load, and obtaining the maximum information coefficient of each influencing factor, includes steps A1-A3:

[0075] A1: Use the mutual information algorithm to calculate the correlation of factors affecting energy consumption;

[0076] A2: Preprocess the maximum mutual information values ​​on different networks to obtain the preprocessed feature matrix. A3: The highest preprocessed value obtained in the feature matrix is ​​the maximum information coefficient.

[0077] In the embodiments of this application, the mutual information algorithm uses the maximum information coefficient (MIC) method.

[0078] Furthermore, the specific implementation methods for steps A1-A3 are as follows:

[0079] The maximum information coefficient (MIC) method was used to screen out the main factors affecting the turbine unit load, and the maximum information coefficient values ​​of each influencing factor were obtained. The training set and test set data were normalized. The correlation calculation and analysis of factors affecting energy consumption, such as water flow, head, reservoir water level, unit load and efficiency, were performed using the maximum information coefficient (MIC) method. The calculation process is as follows:

[0080] In a given hydropower unit dataset (D), input variables (X) and unit energy consumption (Y) are divided into regions (X) and (Y) using grid regions (G). (D|G) represents the distributional change in (D) caused by element pairs within the grid region (G). The maximum mutual information (MI) of the grid on (D) can be defined as the value that maximizes the mutual information between (X) and (Y) within region (G):

[0081]

[0082] Wherein, given a dataset (D) and grid regions (G), (p(a,b)) is the joint probability density of variables (a) and (b), and (p(a)) and (p(b)) are the marginal probability densities of variables (a) and (b) calculated using the histogram estimation method, respectively. MI(D|G) is the mutual information of grid regions (G) on the dataset (D). For a fixed set (D), the normalized MI feature matrix obtained by normalizing the maximum mutual information values ​​on different grids (G) can be defined as follows:

[0083]

[0084] The maximum information coefficient is the highest normalized MI value obtained from the feature matrix, and is defined as:

[0085]

[0086] By calculating the maximum mutual information coefficient, the energy consumption correlation factors of hydropower units with coefficients greater than 0.7 were selected as the optimal combination of influencing factors and used as the model input.

[0087] To reduce the impact of scale differences between features on prediction performance, historical data is normalized. The normalization formula is shown in equation (9):

[0088]

[0089] Where: e represents the historical data before normalization; E represents the historical data after normalization; emax represents the maximum value in the historical data before normalization; and emin represents the minimum value in the historical data before normalization.

[0090] In this embodiment of the application, step S3 uses the preprocessed data as the input data for the first model, and determines the basic hyperparameters of the first model based on the input data. The first model adopts the XGBoost model, and the specific implementation is as follows:

[0091] The normalized data is used as input data for the XGBoost model for further feature selection, and the basic hyperparameters of the XGBoost model are determined based on the input data.

[0092] For the learning rate in the model, the smaller the learning rate, the more trees are needed (because the update step size in each iteration is smaller).

[0093] learning_rate∈[0.01,0.3](0.1 is commonly used).

[0094] n_estimators are automatically determined through early stopping (e.g., stopping when the validation set metric does not improve after 50 consecutive rounds).

[0095] The learning rate is related to the convergence of gradient descent and must meet the following conditions:

[0096] learning_rate × n_estimators ≈ constant

[0097] For the tree depth in the model, the tree depth is (max_depth).

[0098] Initial value: max_depth = 6.

[0099] For small datasets or strong regularization requirements: max_depth∈[3,5].

[0100] For large datasets or complex relationships: max_depth∈[8,12] (requires regularization).

[0101] Formula reference (applicable to balanced trees):

[0102] Maximum number of leaf nodes ≈ 2max_depth

[0103] If the number of samples N is small, the depth should be limited to avoid overfitting.

[0104] max_depth≤log2(N)-1.

[0105] In an optional embodiment, the first model may be a random forest model;

[0106] In another alternative embodiment, the first model may also be a support vector machine.

[0107] In this embodiment of the application, step S4 utilizes an intelligent optimization algorithm to optimize the first model for feature extraction, thereby improving the feature capture capability. The intelligent optimization algorithm used is the Caucasian porcupine optimization algorithm (CPO). The specific implementation method is as follows:

[0108] To address the problem of poor prediction results due to difficulties in hyperparameter tuning during feature selection, the Caucasian Pig Optimization Algorithm (CPO) is used to optimize the XGBoost model for feature extraction, thereby improving the feature capture ability, reducing manual intervention, and thus improving prediction accuracy.

[0109] The core innovation of the CPO algorithm lies in the introduction of cyclic population reduction (CPR) technology. During optimization, CPR accelerates convergence by extracting some CPs from the population, and then reintroducing these individuals into the population to enhance population diversity. Its expression is shown in formula (10), which can be expressed as:

[0110]

[0111] Where: T is the variable that determines the number of iterations; T max N is the maximum number of loops; min The minimum number of individuals in the population; t represents the solution value of the function in the current iteration; N is the population size.

[0112] (1) Global Exploration Phase

[0113] During this phase, based on the defensive behavior of the hooded porcupine, when the predator is at a distance, the hooded porcupine employs both visual and auditory strategies. These strategies involve exploring different areas of the surrounding environment to achieve a comprehensive search.

[0114] First defense strategy: The crowned porcupine raises and vibrates its feathers to warn predators, as shown in formula (11).

[0115]

[0116] Where, x CP For the best CP position, y t A vector generated between the current CP and a random CP:

[0117]

[0118] Second defense strategy: The crowned porcupine enhances its defense capabilities by making loud noises to intimidate potential predators, making them feel threatened and causing them to retreat, as shown in formula (13).

[0119]

[0120] in: It is the position of the i-th individual during iteration; τ1 is the position of the i-th individual in the next iteration; τ2 and τ3 are random numbers that follow a normal distribution; τ2 and τ3 are random numbers between [0,1]. It is the optimal solution for function t; It represents the predator's position at iteration t; It is a binary vector consisting of 0 and 1; r1 and r2 are two random integers between [1, N].

[0121] (2) Partial Development Phase

[0122] The third defense strategy: When a predator approaches, the porcupine will take further multi-layered self-protection measures. Among them, the third defense means secreting a special odor substance. This substance has strong volatility and can spread rapidly around the porcupine, forming an effective "odor barrier" to prevent the enemy from getting close. The mathematical model is shown in formula (14).

[0123]

[0124] Where: x CP For the best CP position, y t This is the vector generated between the current CP and the random CP.

[0125] Fourth defense strategy: When none of the previous strategies are available, the crowned porcupine will launch a physical attack on the predator, and the mathematical model is shown in formula (15).

[0126]

[0127] Where: r3 is a random number between [1, N]; γ t It is a defense factor; S i ′ is an odor diffusion factor; These are parameters that control the search direction; α is the convergence rate; τ4 and τ5 are also random numbers between [0,1]. It is the inelastic collision force generated when an individual physically attacks a predator.

[0128] Optimization steps:

[0129] 1. Initialize the population: Randomly generate multiple sets of XGBoost hyperparameter combinations for tree depth and minimum sample weights of leaf nodes.

[0130] 2. Evaluate fitness: Use cross-validation to evaluate the performance of each hyperparameter combination.

[0131] 3. CPO Optimization Parameters: Utilizing CPO to find the optimal solution within the given range.

[0132] 4. Termination condition: Stop when the maximum number of iterations is reached or the fitness no longer improves.

[0133] 5. Train the final model: Train the XGBoost model with optimal hyperparameters and evaluate its performance.

[0134] In this embodiment of the application, the load prediction of the turbine unit in step S5 and the output of the load prediction result are performed using a combined model of a temporal convolutional network and a bidirectional long short-term memory network (TCN-BiLSTM). The specific implementation method is as follows:

[0135] The depth of the tree and the minimum sample weight of the leaf nodes in the XGBoost model were optimized using CPO. The optimized parameters were assigned to the model, and a combined model of temporal convolutional network and bidirectional long short-term memory network (TCN-BiLSTM) was used to predict the load of the turbine unit. The model was tested using test set data, the load prediction results were output, and various evaluation indicators were calculated to evaluate the model performance.

[0136] The TCN-BiLSTM model, which combines TCN and BiLSTM, leverages the advantages of convolutional networks and bidirectional LSTM to better capture complex dependencies in time series, enhance feature extraction capabilities, and improve prediction accuracy and robustness. The TCN network consists of three layers: a causal convolutional layer, a normalization layer, an activation function, a dropout layer, and residual convolutional connections. The TCN layer processes the input feature sequence through multiple convolutional layers, each with a different dilation rate to capture long-term dependencies. Batch normalization and ReLU activation functions are used to increase model stability and non-linear modeling capabilities. The main formula of the TCN model is as follows: Dilated convolutions increase the receptive field by inserting gaps between elements of the convolutional kernel, as shown in the formula:

[0137]

[0138] Where: s is the element in the sequence; * is the convolution operation; d is the dilation coefficient; K is the convolution kernel size; x s-d·i f(i) is the i-th data in the dilated convolution; f(i) is the i-th element in the convolution kernel.

[0139] Finally, the MaxPooling layer performs dimensionality reduction and feature selection before inputting into the next network. In BiLSTM, this is where:

[0140] 1. Bidirectional LSTM layer: The input sequence is passed to two LSTM networks, one forward (from front to back) and the other backward (from back to front).

[0141] 2. Merging layer: The outputs of two LSTM layers are usually concatenated (connected) or weighted and combined (e.g., addition, averaging, etc.).

[0142] 3. Output Layer: Typically a fully connected layer, used for prediction in classification, regression, or other tasks. Features extracted by the TCN are input into the LSTM unit, which is configured as a bidirectional LSTM. This ensures that the input at each time step is influenced not only by previous time steps but also by information from subsequent time steps. BatchNormalization standardizes the output of the BiLSTM, making the model more stable. Finally, the output layer uses a fully connected layer to predict the target variable (energy consumption of the hydroelectric turbine). The BiLSTM calculation formula is as follows: The LSTM layer consists of gating units, mainly composed of a forget gate, an input gate, and an output gate. The forget gate forgets information from the previous time step's unit state, the input gate updates the unit state, and the output gate determines the hidden state vector for the next time step. The LSTM expression is:

[0143]

[0144] Among them: W f W i W o W c and b f b i b o b c These are the weight matrices and bias terms for each module; h t-1 The hidden state vector from the previous time step; the input x at the current time step. t σ is the Sigmoid activation function; tanh is the hyperbolic tangent activation function; and is the Hadamard product.

[0145] The expression for BiLSTM is:

[0146] h t =[h f ,h b ] T

[0147] Where: h f h is the hidden layer output of the feedforward LSTM layer at time t; b The hidden layer output of the backward LSTM layer at time t; h t This is the result of combining the forward LSTM layer and the backward LSTM layer.

[0148] In an optional embodiment, the combined model employs a multilayer perceptron-long short-term memory network.

[0149] In another alternative embodiment, the combined model employs a multilayer perceptron-gated recurrent unit.

[0150] Further evaluation indicators:

[0151] Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²) are calculated using formulas (18), (19), (20), and (21):

[0152]

[0153]

[0154] Where: y t and y t ' and ' represent the actual and predicted load power values ​​at time t, respectively; This represents the average value of the actual load power; N is the number of samples used for testing.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0156] Example 3 illustrates a load forecasting method combining the maximum information coefficient method and deep learning. It should be noted that the technical solution of this load forecasting system combining the maximum information coefficient method and deep learning is based on the same concept as the load forecasting method combining the maximum information coefficient method and deep learning described above. Details not described in detail in this example can be found in the description of the load forecasting method combining the maximum information coefficient method and deep learning described above.

[0157] This embodiment also provides a load forecasting system combining the maximum information coefficient method and deep learning, including:

[0158] The screening module uses the maximum information coefficient (MIC) method to screen out the main factors affecting the turbine unit load, obtains the maximum information coefficient value of each influencing factor, and normalizes the training set and test set data.

[0159] The module determines the basic hyperparameters of the XGBoost model by using the normalized data as input data for further feature selection.

[0160] The optimization module addresses the problem of poor prediction results caused by difficulties in hyperparameter tuning during feature selection. It utilizes the Crowned Hog Optimization (CPO) algorithm to optimize the XGBoost model for feature extraction, thereby improving the feature capture ability, reducing manual intervention, and thus improving prediction accuracy.

[0161] The prediction module optimizes the tree depth and minimum sample weights of the leaf nodes in the XGBoost model using CPO, assigns the optimized parameters to the model, and uses a combined model of temporal convolutional network and bidirectional long short-term memory network (TCN-BiLSTM) to predict the turbine unit load. The model is tested using test set data, the load prediction results are output, and various evaluation indicators are calculated to evaluate the model performance.

[0162] This embodiment also provides an electronic device suitable for load forecasting that combines the maximum information coefficient method and deep learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the load forecasting method that combines the maximum information coefficient method and deep learning as proposed in the above embodiment.

[0163] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the load prediction method combining the maximum information coefficient method and deep learning as proposed in the above embodiments.

[0164] The storage medium proposed in this embodiment and the load prediction method combining the maximum information coefficient method and deep learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0165] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0167] Example 4, Reference Figures 2-4 This invention provides a simulation experiment of a load prediction method combining the maximum information coefficient method and deep learning. The data processed by the isolated forest is divided into training and test sets at a ratio of 80% and 20%, respectively. The training set contains 8000 data points with 8 variables, and the test set contains 2000 data points with 8 variables. The training set is used to train the deep learning model and adjust its structural parameters, while the test set is used to evaluate the predictive performance of the trained deep learning model.

[0168] Table 1. Experimental Environment

[0169]

[0170]

[0171] Table 2. Overall Model Parameter Settings

[0172]

[0173] Table 3. Calculation results of MIC coefficient values

[0174]

[0175]

[0176] To verify the superiority of the Crowned Porcupine Optimization (CPO) algorithm, it was compared with current mainstream optimization algorithms, including Gradient Optimization (GBO), Grusonite Optimization (GTO), African Vulture Optimization (AVOA), Whale Optimization (WOA), and Grey Wolf Optimization (GWO). Performance on optimization tasks was evaluated by analyzing fitness, convergence speed, and standard deviation. The experimental results are as follows: Figure 2 and Figure 3 As shown.

[0177] from Figure 2 A comparison of fitness and convergence among different optimization algorithms reveals that CPO converges rapidly in the early stages of iteration, with its fitness value dropping sharply from a relatively high initial level and stabilizing at a value close to -1.6 after 5 iterations, demonstrating its powerful global search capabilities and extremely fast convergence speed. In contrast, other algorithms such as GBO and GTO, while exhibiting a gradual decrease in fitness value, have slower convergence speeds and ultimately higher stable fitness values. WOA, AVOA, and GWO all show significant fluctuations during optimization and slow convergence speeds, with their fitness values ​​eventually stabilizing at relatively high levels, indicating that these algorithms are prone to getting trapped in local optima and lack sufficient robustness. WOA, in particular, while showing some decrease in fitness value in the early stages, experiences significant fluctuations in fitness value in the later stages and fails to achieve stable convergence.

[0178] Standard deviation is a commonly used metric when comparing different optimization algorithms. A lower standard deviation usually means that the algorithm not only finds better solutions but also exhibits more consistent and stable performance. To more intuitively demonstrate the effectiveness of different optimization algorithms, in Figure 3 The standard deviation comparison results of various algorithms in the mode decomposition optimization of hydrogen are presented. From Figure 3Observations show that CPO has the lowest standard deviation at 0.1434, demonstrating its strong accuracy and stability across multiple optimization tasks. CPO excels in accuracy, convergence speed, and global search capability, finding near-global optimal solutions in a short time and exhibiting strong adaptability, particularly in complex and dynamic optimization environments.

[0179] After feature extraction using the CPO-XGBoost model, the data is then input into the TCN-BiLSTM model for load prediction. The comparison results with other models are as follows: Figure 4 As shown in Table 4.

[0180] Table 4. Comparison Results of Indicators for Different Models

[0181]

[0182] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A load forecasting method combining the maximum information coefficient method and deep learning, characterized in that: include: The main factors affecting the load of the turbine unit were screened, and the maximum information coefficient of each influencing factor was obtained. The preprocessed data is used as the input data for the first model, and the basic hyperparameters of the first model are determined based on the input data. The first model is optimized using intelligent optimization algorithms to improve feature extraction capabilities; The load of the turbine unit is predicted, and the load prediction results are output.

2. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 1, characterized in that, The process of screening the main factors affecting the turbine unit load yields the maximum information coefficient for each influencing factor, including: The mutual information algorithm is used to calculate the correlation of factors affecting energy consumption; The maximum mutual information values ​​on different networks are preprocessed to obtain the preprocessed feature matrix. The highest preprocessed value obtained from the feature matrix is ​​the maximum information coefficient.

3. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 2, characterized in that, The first model is optimized using an intelligent optimization algorithm to improve feature extraction and enhance features. Scraping capabilities, including: The first model hyperparameter combination is generated by randomly generating multiple sets of tree depths and minimum sample weights for leaf nodes; Use cross-validation to evaluate the performance of each hyperparameter combination; The optimal solution within the range is found using intelligent optimization algorithms; Stop when the maximum number of iterations is reached or the fitness no longer improves; The first model was trained with optimal hyperparameters and its performance was evaluated.

4. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 3, characterized in that, The prediction of turbine unit load and the output of load prediction results include: The optimized parameters are assigned to the first model, and the combined model is used to predict the turbine unit load. The model is tested using test set data, the load forecast results are output, and the evaluation index is calculated to assess the model performance.

5. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 4, characterized in that, The intelligent optimization algorithm incorporates a cyclic population reduction technique. During the optimization process, this technique accelerates convergence by extracting a portion of the population's core populations (CPs) and then reintroducing individuals into the population. This can be expressed as: Where T is the variable that determines the number of iterations; T max N is the maximum number of loops; min The minimum number of individuals in the population; t represents the solution value of the function in the current iteration; N is the population size.

6. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 5, characterized in that, The correlation calculation of factors affecting energy consumption using the mutual information algorithm is expressed as follows: Wherein, given a dataset D and grid regions G, p(a,b) is the joint probability density of variables a and b, p(a) and p(b) are the marginal probability densities of variables a and b calculated by the histogram estimation method, respectively, and MI(D|G) is the mutual information of grid regions G on dataset D. For a fixed given dataset D, the normalized MI feature matrix is ​​obtained by normalizing the maximum mutual information values ​​on different grids G as follows: The maximum information coefficient is the highest normalized MI value obtained from the feature matrix, expressed as: By calculating the maximum mutual information coefficient, the energy consumption correlation factors of hydropower units with coefficients greater than the threshold are selected as the influencing factor combination sample set as the model input. The historical data is normalized and represented as follows: Where e represents the historical data before normalization; E represents the historical data after normalization; e max e represents the maximum value in the historical data before normalization; min This represents the minimum value in the historical data before normalization.

7. The load forecasting method combining the maximum information coefficient method and deep learning as described in claim 6, characterized in that, The computational evaluation index assesses the model performance and is expressed as follows: Among them, y t and y t ' and ' represent the actual and predicted load power values ​​at time t, respectively; This represents the average value of the actual load power; N is the number of samples used for testing; MSE represents the mean square error; RMSE represents the root mean square error; MAPE represents the mean absolute percentage error; R 2 It is expressed as the coefficient of determination.

8. A load forecasting system combining the maximum information coefficient method and deep learning, applying the method described in any one of claims 1-7, characterized in that, include: The module includes a screening module, a determination module, an optimization module, and a prediction module.

9. An electronic device, comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the load prediction method combining the maximum information coefficient method and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the load forecasting method combining the maximum information coefficient method and deep learning as described in any one of claims 1 to 7.

Images