Icebreaker power system load prediction method based on SAO-BiLSTM

Abstract

The invention provides an icebreaker power system load prediction method based on SAO-BiLSTM. The method comprises the steps of collecting environmental factor data of an icebreaker to be subjected to load prediction; decomposing the collected environmental factor data to obtain corresponding IMF components and residual components, and forming a feature sequence; selecting a principal component of which the cumulative contribution rate is greater than a set value from the feature sequence, and projecting the environmental factor data onto the selected principal component to obtain dimension-reduced data; inputting the data after dimension reduction into an icebreaker power system load prediction model to obtain a load prediction value; the icebreaker power system load prediction model is obtained by optimizing network parameters of a BiLSTM network by adopting an improved snow ablation optimization algorithm; the improved snow ablation optimization algorithm comprises the steps of adding a heat transfer strategy in an exploration stage and adding a condensation strategy in a development stage.

Description

Technical Field

[0001] This invention belongs to the field of ship power system load forecasting technology, specifically relating to a load forecasting method for icebreaker power systems based on SAO-BiLSTM. Background Technology

[0002] Icebreakers operate in polar environments, and their electrical systems not only need to withstand the challenges of extreme cold but also meet complex and variable operational requirements. During icebreaking operations, especially under repeated impact conditions, icebreakers continuously break through thick ice layers by accelerating and reversing. Each impact causes a sudden increase in ice load (the interaction between ice and structures), leading to a rapid increase in torque load. Subsequently, the torque load decreases sharply when the ship reverses, and then increases dramatically upon the next impact. This impact-load fluctuation places high demands on the stability of the electric propulsion system, potentially causing voltage fluctuations and frequency instability, which in turn leads to decreased system efficiency, affecting propulsion efficiency and system reliability.

[0003] Accurate load forecasting technology can effectively optimize energy allocation, improve fuel efficiency, reduce power loss caused by load imbalance, help optimize generator power supply strategies, and effectively mitigate the impact of drastic load fluctuations on the power grid, thereby improving power system stability and ensuring the safety of mission execution. Before load forecasting, data processing can significantly improve the accuracy and reliability of the forecasting model by extracting useful features. The commonly used "decomposition-forecasting" method effectively explores the patterns of power system load, but it has limitations due to its own network structure when dealing with load fluctuations caused by constantly changing propulsion load torque and torque variations applied to propeller blades during icebreaking operations. Machine learning methods are widely used in the field of power load forecasting, including support vector regression, radial basis function neural networks, and long short-term memory neural networks; however, machine learning methods generally suffer from difficulties in hyperparameter selection and low computational efficiency. Summary of the Invention

[0004] Purpose of the invention: To address the problem that existing ship power load forecasting models are difficult to adapt to the power load forecasting requirements of icebreakers under extreme working environments and conditions, this invention proposes a power system load forecasting method for icebreakers based on SAO-BiLSTM.

[0005] Technical solution: A load forecasting method for icebreaker power systems based on SAO-BiLSTM, comprising:

[0006] Step 1: Collect environmental factor data of the icebreaker to be predicted for load forecasting; decompose the collected environmental factor data into EEMD to obtain the corresponding IMF components and residual components, which constitute a feature sequence; use DS-KPCA to select principal components with a cumulative contribution rate greater than 95% from the feature sequence, and project the environmental factor data onto the selected principal components to obtain the dimensionality-reduced data;

[0007] Step 2: Input the dimensionality-reduced data into the icebreaker power system load forecasting model to obtain the load forecast values;

[0008] The icebreaker power system load prediction model is obtained by optimizing the network parameters of the BiLSTM network using an improved snow ablation optimization algorithm. The optimization process includes an initialization phase, an exploration phase, and a development phase. The improved snow ablation optimization algorithm includes adding a heat transfer strategy in the exploration phase and a condensation strategy in the development phase. In the heat transfer strategy, individuals are sorted by fitness from low to high, and high-temperature water vapor is defined as the top five individuals in terms of population fitness value, while the remaining individuals are low-temperature water vapor. Excellent individuals guide poor individuals to explore. In the condensation strategy, the number of subpopulations is increased in the development phase and decreased in the exploration phase.

[0009] Furthermore, the network parameters of the BiLSTM network include: the number of hidden units, the training period, and the learning rate.

[0010] Furthermore, an improved snow ablation optimization algorithm is used to optimize the network parameters of the BiLSTM network. Specific operations include:

[0011] Step 1: Initialize the snow ablation optimization algorithm parameters as the initial population. Randomly divide the entire initial population into two equal-sized subpopulations, including the development subpopulation N. b and exploring subgroup N a ;

[0012] Step 2: Calculate the fitness function value of each individual in the initial population, sort the fitness function values, and find the best individual;

[0013] Step 3: Let t=0, i=1;

[0014] Step 4: Execute the heat transfer strategy and update the position of each component;

[0015] Step 5: Calculate the fitness function value of the updated individual position and determine whether it is better than the current best individual, t=t+1, i=i+1;

[0016] Step 6: Determine if i is greater than the explored subpopulation N a If it is greater than , calculate the snow melting coefficient and snow melting rate, and proceed to step 7; if it is not greater than , proceed to step 4.

[0017] Step 7: Let i = 1;

[0018] Step 8: Execute the heat transfer strategy and update the position of each element;

[0019] Step 9: Execute the condensation strategy and update the positions of each body;

[0020] Step 10: Determine if i is greater than the development subgroup N. b If i is greater than the development subgroup N b If T is greater than 0.4Tmax, then determine if i=1 and proceed to step 11; otherwise, proceed to step 13; if i is not greater than the development subgroup N... b If i = i + 1, then proceed to step 8;

[0021] Step 11: Execute the condensation strategy and update the positions of each body;

[0022] Step 12: Determine if i is greater than N. If it is, calculate the fitness function value, T=t+1, and proceed to step 13; otherwise, i=i+1 and proceed to step 11. N represents the population size, specifically the total number of snow particles / water vapor molecules in the algorithm.

[0023] Step 13: Determine if T is greater than Tmax. If it is, the algorithm ends and outputs the optimal solution for the network parameters of the BiLSTM network. Otherwise, proceed to step 4.

[0024] Furthermore, the execution of the heat transfer strategy and updating the positions of each body includes:

[0025] First sub-process: Update the position of the low-temperature water vapor individual using the following formula:

[0026]

[0027]

[0028]

[0029] In the formula, This refers to an individual randomly selected from the set of the five individuals with the highest fitness values ​​in the population; , It is a random number uniformly distributed in the range [0,1]. , It is a random number between [0,1]. AF represents the heat transfer control factor and the heat decay factor.

[0030] The second sub-process involves mutual learning and position updates among high-temperature water vapor individuals to improve the quality of superior individuals and explore more valuable solution space regions. The positions of the high-temperature water vapor individuals are updated using the following formula:

[0031]

[0032] in, Let i be the position of the (i-1)th individual. Let i be the position of the (i+1)th individual. Represents a random number between [0,1]. This represents the positional difference between the current high-performing individual and other low-performing individuals. For the current individual, It is one of the individuals whose fitness value is in the middle of the entire population.

[0033] Furthermore, the exploration process, which updates the positions of each entity, is represented as follows:

[0034]

[0035]

[0036]

[0037] in, For the updated position of the individual, Indicates the current position of the individual. This represents a random number vector based on Brownian motion. It is a random number. This is the current best solution, where and These represent the second and third best individuals in the current population, respectively. For individuals randomly selected from an elite set, The central location of the population; N1 represents the centroid position of individuals with fitness values ​​in the top 50%, and N1 is the number of leaders, which is half the number of the cluster.

[0038] Furthermore, the snow ablation coefficient is calculated using the following formula:

[0039]

[0040] Where t is the current iteration number, It is the maximum value;

[0041] The snow melt rate is calculated using the following formula:

[0042]

[0043] In the formula, T represents the daily average temperature.

[0044] Furthermore, the development process described above, which updates the positions of each entity, is represented as follows:

[0045]

[0046] in, for Random numbers.

[0047] Furthermore, the execution of the condensation strategy, updating the positions of each entity, includes:

[0048]

[0049]

[0050]

[0051] in, Let be the position of the random individual in the population responsible for the t-th iteration, and let e increase as the number of iterations increases. It is a random number between [−1, 1].

[0052] Furthermore, the step of selecting principal components with a cumulative contribution rate greater than a set value from the feature sequence, projecting the environmental factor data onto the selected principal components to obtain dimensionality-reduced data, is specifically implemented using dynamic sparse kernel principal component analysis, including the following steps:

[0053] Based on the environmental factor data of the icebreaker collected for load forecasting, the real-time operating intensity index is calculated:

[0054]

[0055] Where IceThick(t) is the sea ice thickness, Wind(t) is the wind speed at 10 meters, and dv / dt is the icebreaker's acceleration. , , The corresponding weights;

[0056] Based on real-time operating conditions and intensity indicators, the feature sparsity is adaptively adjusted:

[0057]

[0058] in, The sparsity is the magnitude. To minimize sparsity, The maximum sparsity is given by k, which is the adjustment slope.

[0059] Construct an objective function with the dimensionality-reduced projection matrix W as the optimization variable, and achieve automatic selection of principal components through sparse constraints:

[0060]

[0061] Among them, regularity strength Dynamically adjusts with sparsity. For kernel mapping functions, For characteristic sequences;

[0062] The objective function is solved using the proximal gradient descent method to obtain the optimal dynamic sparse projection matrix W(t);

[0063] Projecting the feature sequence onto the optimal dynamic sparse projection matrix W(t) yields the dimensionality-reduced data Z(t):

[0064] .

[0065] Beneficial Effects: This invention improves load forecasting accuracy through the combined application of data decomposition, feature extraction, and optimization algorithms, especially when dealing with icebreaking conditions, showing a significant improvement over traditional forecasting models. This model provides effective data support for the safe and stable operation of the icebreaker's power system and unit operation strategies, improving system reliability and operational efficiency. Compared with existing technologies, this invention has the following advantages:

[0066] (1) The method of this invention constructs an EEMD-KPAC-BiLSTM network for load forecasting of icebreaker power systems. EEMD performs multi-scale decomposition of time series, which can better capture the load fluctuation characteristics of different frequencies and improve the feature extraction capability. DS-KPCA effectively reduces the data dimension and retains key nonlinear features, enhancing the model's ability to identify complex patterns. Compared with general BiLSTM models, the error index is reduced, effectively solving the time lag problem.

[0067] (2) The method of the present invention adopts the dynamic sparsity method based on the navigation environment parameters of icebreakers to improve the kernel principal component analysis method. This method couples environmental perception with sparsity control and realizes intelligent feature extraction with dimensionality reduction according to computational needs. While ensuring the prediction accuracy of thick ice layer areas, it significantly improves the computational efficiency of open water areas.

[0068] (3) The present invention adopts a heat transfer-condensation dual strategy to improve SAO. This strategy enhances the global search capability and post-development capability of the algorithm, thereby significantly improving the efficiency and convergence speed of BiLSTM network parameter optimization, enabling it to effectively cope with the load fluctuation of icebreakers in extreme climates and environments, solving the high-frequency load noise interference and prediction lag problem in the power system of polar ships, and further improving the prediction accuracy and system stability. Attached Figure Description

[0069] Figure 1 This is a diagram of a load prediction model for an icebreaker power system based on a SAO-BiLSTM neural network proposed in this invention.

[0070] Figure 2 A graph showing the actual load data for icebreaker navigation;

[0071] Figure 3 This is a flowchart of the ISAO algorithm.

[0072] Figure 4 This is a comparison chart of the predictive model performance. Detailed Implementation

[0073] The technical features of the present invention will be further described clearly and completely below with reference to the accompanying drawings. Obviously, the described examples are only some embodiments of the present invention, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0074] This invention proposes a load forecasting method for icebreaker power systems based on a SAO-BiLSTM neural network. This method effectively addresses high-frequency load fluctuations in power systems during icebreaking operations. By introducing EEMD and DS-KPCA data preprocessing techniques, the quality of input data is significantly improved, and noise interference is reduced. Based on a dual heat transfer-condensation strategy, the method overcomes the limitations of SAO in later development and its susceptibility to local optima. It further enhances the efficiency and convergence speed of BiLSTM network parameter optimization, effectively addressing load fluctuations in icebreakers under extreme weather conditions. This solves the problems of high-frequency load noise interference and prediction lag in polar vessel power systems, providing a technical solution for dynamic energy scheduling and system reliability management in polar environments. Figure 1 As shown, the specific steps include:

[0075] Step 1: Collect navigation environment parameters and historical load records for an icebreaker. These navigation environment parameters include environmental parameters and weather parameters. In this embodiment of the invention, the load recording period is from June 5th to June 21st, 2023. The data collection frequency is once per minute, with 1440 sampling points per day, resulting in a total of 23040 load values. The original load data is as follows: Figure 2 As shown.

[0076] Each navigation environment parameter is decomposed using EEMD (Ensemble Empirical Mode Decomposition) to obtain the corresponding number of IMF components and residual components, forming a feature sequence. Empirical Mode Decomposition, as an adaptive method, can decompose complex signals into several Intrinsic Mode Functions (IMFs). However, traditional EMD suffers from mode aliasing. Therefore, this embodiment of the invention uses EEMD to introduce white noise and decompose multiple times to obtain IMF components and residual components, forming a feature sequence. In this embodiment, eight typical environmental data types are selected, as shown in Table 1.

[0077] Table 1 Components of navigation environment parameters

[0078]

[0079] Step 2: Assemble the obtained N rows and M columns of feature sequences into matrix X. Perform zero-mean normalization on each row of matrix X, and calculate the eigenvalues ​​and eigenvectors of the covariance matrix. Arrange the eigenvectors into a matrix based on their corresponding eigenvalues ​​from top to bottom. Select principal components with a cumulative contribution rate greater than 95% to ensure excellent representativeness, accurately reflect the key information of the original feature sequences, and possess stronger comprehensive information carrying capacity. The implementation process is shown below:

[0080] The real-time operating condition intensity index is calculated based on the navigation environment parameters of the icebreaker, with a range of [0,1]. The calculation method is as follows:

[0081]

[0082] Where IceThick(t) is the sea ice thickness, Wind(t) is the wind speed at 10 meters, and dv / dt is the icebreaker's acceleration. Take 0.6, Take 0.2, Take 0.2.

[0083] The DS-KPCA (Dynamic Sparse-Kernel Principal Components Analysis) method based on icebreaker navigation environment parameters is adopted to map the feature sequence obtained in step 1 from the original space to a high-dimensional feature space. The sparsity is adjusted according to the operational intensity index to perform dimensionality reduction as needed. This significantly improves the computational efficiency in open waters while maintaining prediction accuracy in thick ice areas. The feature sparsity is adaptively adjusted according to the operational intensity index.

[0084]

[0085] in, The sparsity is the magnitude. To minimize sparsity, is the maximum sparsity, and k is the adjustment slope.

[0086] Construct an objective function with the dimensionality-reduced projection matrix W as the optimization variable, and achieve automatic selection of principal components through sparse constraints:

[0087]

[0088] Among them, regularity strength Dynamically adjusts with sparsity. For kernel mapping functions, It is a characteristic sequence.

[0089] The above optimization problem is solved iteratively using the proximal gradient descent method, and the optimal dynamic sparse projection matrix W(t) is finally obtained. The sparsity mode of this matrix is ​​dynamically determined by the sparsity s(t) under the current working condition intensity.

[0090] Subsequently, by performing the multiplication of the kernel matrix and the projection matrix, the dimensionality-reduced data Z(t) is directly obtained:

[0091] .

[0092] Thus, the dynamic selection of principal components and data projection were completed simultaneously using the DS-KPCA method. The obtained Z(t) is the dimensionality-reduced feature data, which retains the information of the original data on the key principal components, while eliminating redundant features and adapting to the current navigation conditions of the icebreaker. This Z(t) will be used as the input for the subsequent BiLSTM load prediction model.

[0093] Step 3: Construct a load forecasting model for the icebreaker's power system using a BiLSTM neural network improved with the ISAO algorithm. Specific operations include:

[0094] The BiLSTM neural network makes a preliminary prediction of the load based on historical load data and corresponding historical navigation environment parameters.

[0095] The BiLSTM neural network will now be explained in more detail.

[0096] The output of the LSTM network is represented as follows:

[0097]

[0098] in , , These represent the activation values ​​of the input gate, forget gate, and output gate, respectively. This represents a candidate value representing the current cell state. Indicates the current cell state. Indicates the output of the current cell. This represents the sigmoid function. W is the weighting coefficient, and b is the corresponding bias.

[0099] The BiLSTM neural network, while maintaining the original internal structure of the Long Short-Term Memory (LSTM) unit, adjusts the network structure on the hidden layers to include forward LSTM layers and backward LSTM layers. Its mathematical expression is as follows:

[0100]

[0101]

[0102]

[0103] In the formula, Indicates the state of the forward hidden layer; This represents the state of the backward hidden layer; T is the length of the input sequence.

[0104] BiLSTM neural network models exhibit low prediction accuracy, high error rates, and significant time lag, failing to accurately capture the dynamic characteristics of ship power load changes. Furthermore, they are insensitive to load abrupt changes and nonlinear characteristics, leading to substantial discrepancies between predicted and actual load data. Therefore, this invention employs an improved SAO optimization algorithm to train the BiLSTM network. This involves initializing parameters, constructing a fitness function, calculating fitness, and updating individual positions to output optimal parameters. Specifically, the improved SAO algorithm optimizes the number of hidden units, training period, and learning rate of the BiLSTM neural network to obtain the optimal solution. The optimization process mainly includes three stages: initialization, exploration, and development.

[0105] The improved SAO algorithm is used to optimize the number of hidden units, training cycle, and learning rate of the BiLSTM neural network, and the optimal solution is explained below, including:

[0106] An initial population is randomly generated, which represents the parameters of the snow ablation optimization algorithm, and is expressed as follows:

[0107]

[0108] Where L and U represent the lower and upper bounds of the solution space, respectively. It is in A number randomly generated between [variables].

[0109] The entire population is randomly divided into two equal-sized subpopulations, one for exploration and the other for development. As iterations proceed, the development subpopulation N... b The size of the subgroup N gradually decreases, while the exploration subgroup N... a The scale has expanded accordingly.

[0110] The exploration process was established by considering Brownian motion to simulate the highly dispersed characteristics of liquid water evaporating into vapor during snow melting, and its position update formula is as follows:

[0111]

[0112]

[0113]

[0114] in, For the updated position of the individual, Indicates the current position of the individual. This represents a random number vector based on Brownian motion. It is a random number. This is the current best solution, where and These represent the second and third best individuals in the current population, respectively. For individuals randomly selected from an elite set, This is the central location of the population. N1 represents the centroid position of individuals with fitness values ​​in the top 50%, and N1 is the number of leaders, which is half the number of the cluster.

[0115] During the development phase of the SAO algorithm, its mathematical model was established based on the melting behavior of snow into water:

[0116]

[0117] Where M represents the snow melt rate, DDF represents the snow ablation coefficient, ranging from 0.35 to 0.6, and T is the daily average temperature.

[0118] The trend of DDF changing over time is as follows:

[0119]

[0120] Where t is the current iteration number, This is the maximum value. In the SAO algorithm, the snow melt rate is calculated using the following formula:

[0121]

[0122] Its update position formula is:

[0123]

[0124] in, for Random numbers.

[0125] Due to changes in ambient temperature, snow gradually melts from a solid to a liquid, and then evaporates from a liquid to a gas. Snow can also sublimate from a solid to a gas when it absorbs a large amount of heat in a short period. However, in the later stages of the algorithm, since all the snow has been converted to gas, the algorithm's development capability decreases, and the quality of the solution deteriorates. To address this problem, inspired by natural heat transfer and condensation phenomena, an improved SAO algorithm is proposed, which enhances the snow melting optimization algorithm with heat transfer and condensation strategies. When gases interact, a heat transfer strategy is added, where the direction of gas movement is from low temperature to high temperature. This strategy simulates high-temperature gas, guiding low-temperature gas towards it through heat transfer, thereby improving the algorithm's global exploration capability. Then, a condensation strategy is added in the development stage to simulate condensation, converting some gas into liquid, improving the algorithm's later development capability, accelerating the convergence speed, and achieving a balance between exploration and development.

[0126] (1) Heat transfer strategy

[0127] In nature, due to temperature differences between gases, heat is transferred from high to low altitudes. This phenomenon alters the direction of motion of the colder gas, causing it to move closer to the hotter gas. In a high-temperature environment, the kinetic energy between gases increases, leading to collisions and mutual influence. This strategy simulates the process of hot water vapor transferring heat to cold water vapor, the movement of colder water vapor towards hotter water vapor, and the mutual influence of hot water vapor through heat transfer. In this strategy, individuals are ranked from low to high fitness. Hot water vapor is defined as the top five individuals in terms of fitness value. The remaining individuals are considered cold water vapor. Superior individuals guide weaker individuals in exploration, which helps improve the algorithm's optimization ability. This strategy divides the heat transfer process into two sub-processes:

[0128] The first sub-process involves heat transfer from high-temperature water vapor to low-temperature water vapor. It's necessary to determine whether all low-temperature water vapor moves towards the high-temperature direction. Because gases undergo random motion, it cannot be guaranteed that all water vapor will move from low to high temperature. Therefore, a restart method is needed to correct the direction of gas movement. The position of individual low-temperature water vapor particles is updated using the following formula:

[0129]

[0130]

[0131]

[0132] In the formula, This refers to an individual randomly selected from the set of the five individuals with the highest fitness values ​​in the population; , It is a random number uniformly distributed in the range [0,1]. , It is a random number between [0,1]. The heat transfer control factor is set to 0.3. AF is the heat decay factor, which gradually approaches 0.01 as the number of iterations increases.

[0133] The second sub-process demonstrates that heat is also transferred and influenced between individuals within the high-temperature water vapor system. This mutual influence enhances the qualities of high-performing individuals, facilitating their exploration of more valuable fields. The specific formula is as follows:

[0134]

[0135] in, Let i be the position of the (i-1)th individual. Let i be the position of the (i+1)th individual. Represents a random number between [0,1]. This represents the positional difference between the current top performer and other less performer individuals. For the current individual, It is one of the individuals whose fitness value is in the middle of the entire population.

[0136] (2) Coagulation strategy

[0137] Since nature includes not only sublimation, melting, and evaporation, but also condensation and solidification, these five phenomena combine to form a thermal cycle. The original SAO algorithm suffers from insufficient late-stage development capabilities, hence this strategy is proposed. Inspired by condensation, this strategy simulates the condensation process. Because sublimation, melting, and evaporation are all endothermic processes, they lead to a decrease in ambient temperature, causing some water vapor to release heat and condense back into water, thus maintaining the thermal cycle. In the SAO algorithm, this process manifests as an increase in the number of subpopulations responsible for mining and a decrease in the number of subpopulations responsible for exploration, improving the algorithm's late-stage mining capabilities, convergence speed, and solution quality. The specific formula is as follows:

[0138]

[0139]

[0140]

[0141] in, Let be the position of the random individual in the population responsible for the t-th iteration, and let e increase as the number of iterations increases. It is a random number between [−1, 1].

[0142] The above process is as follows Figure 3 As shown.

[0143] The first 80% of historical load data was used as the model training sample, and the last 20% was used as the test sample. These samples were then compared with the actual load data. The predictive model performance is as follows: Figure 4 As shown.

[0144] Step 4: Select the root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R²), Person correlation coefficient (r), and mean absolute error (MAE) as the performance evaluation indicators for the load forecasting model. The root mean square error function is selected as the fitness function for the SAO algorithm.

[0145] The calculation formula is as follows:

[0146]

[0147]

[0148]

[0149]

[0150]

[0151] In the formula, N is the total number of test samples. To test the actual values ​​of the samples in the test set, The predicted value is obtained based on the prediction model in this paper. The average value of the samples in the test set. The average value obtained for the prediction model.

[0152] The model performance was evaluated based on the predictive performance evaluation index. For specific data comparison, please refer to Table 2.

[0153] Table 2 Comparison of data for each evaluation indicator in the prediction results.

[0154] .

Claims

1. A SAO-BiLSTM-based load forecasting method for an icebreaker power system, characterized in that: The method comprises the following steps: Step 1: collecting environmental factor data of an icebreaker to be subjected to load prediction; The collected environmental factor data is subjected to EEMD decomposition to obtain corresponding IMF components and residual components, thereby forming a characteristic sequence; DS-KPCA is used to select principal components with cumulative contribution rate greater than 95% from the characteristic sequence, and the environmental factor data is projected onto the selected principal components, thereby obtaining dimension-reduced data; Step 2: inputting the dimension-reduced data into an icebreaker power system load prediction model to obtain a load prediction value; The icebreaker power system load prediction model is obtained by optimizing network parameters of a BiLSTM network by using an improved snow ablation optimization algorithm; the optimization process comprises an initialization stage, an exploration stage and a development stage; the improved snow ablation optimization algorithm comprises adding a heat transfer strategy in the exploration stage and adding a condensation strategy in the development stage; in the heat transfer strategy, individuals are sorted from low to high according to fitness, high-temperature water vapor is defined as individuals with top five fitness values in the population, and the rest of the individuals are low-temperature water vapor, and excellent individuals guide poor individuals to explore; in the condensation strategy, the number of subpopulations in the development stage is increased, and the number of subpopulations in the exploration stage is reduced.

2. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 1, characterized in that: The network parameters of the BiLSTM network comprise a number of hidden units, a training period and a learning rate.

3. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 2, characterized in that: The network parameters of the BiLSTM network are optimized by using the improved snow ablation optimization algorithm, and the specific operation comprises the following steps: Step 1: initialize the parameters of the snow melt optimization algorithm as the initial population, randomly divide the entire initial population into two equal-sized subpopulations, including the development subpopulation N b and the exploration subpopulation N a ; Step 2: calculating fitness function values of each individual in the initial population, sorting the fitness function values, and finding the best individual; Step 3: setting t=0 and i=1; Step 4: executing the heat transfer strategy to update the positions of the individuals; in the heat transfer strategy, individuals are sorted from low to high according to fitness, high-temperature water vapor is defined as individuals with top five fitness values in the population, and the rest of the individuals are low-temperature water vapor, and excellent individuals guide poor individuals to explore; Step 5: calculating the fitness function values of the updated individual positions, determining whether they are better than the current best individual, setting t=t+1 and i=i+1; Step 6: Determine if i is greater than the exploration subpopulation N a If yes, calculate the snowmelt coefficient and the snowmelt rate and perform step 7; if no, perform step 4; Step 7: setting i=1; Step 8: executing the heat transfer strategy to update the positions of the individuals; Step 9: executing the condensation strategy to update the positions of the individuals; in the condensation strategy, the number of subpopulations in the development stage is increased, and the number of subpopulations in the exploration stage is reduced; Step 10: Determine if i is greater than development subpopulation N b , if i is greater than development subpopulation N b , determine if T is greater than 0.4Tmax, if yes, then i = 1, perform step 11; otherwise, perform step 13; if i is not greater than development subpopulation N b , then i = i + 1, perform step 8; Step 11: executing the condensation strategy to update the positions of the individuals; Step 12: determining whether i is greater than N, if yes, calculating the fitness function value, setting T=t+1 and executing step 13; otherwise, setting i=i+1 and executing step 11; N represents the population size; Step 13: determining whether T is greater than Tmax, if yes, the algorithm ends, and the optimal solution of the network parameters of the BiLSTM network is output; otherwise, step 4 is executed.

4. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 2, characterized in that: The heat transfer strategy is executed to update the positions of the individuals, and the method comprises the following steps: A first sub-process: the positions of the low-temperature water vapor individuals are updated by using the following formula: In the formula, represents an individual randomly selected from a set of 5 individuals with the highest fitness values in the population; , is a random number uniformly distributed in the range [0, 1]; , is a random number between [0, 1]; is a heat transfer control factor, and AF represents a heat attenuation factor; A second sub-process: mutual learning and position updating are performed among the high-temperature water vapor individuals to improve the quality of the excellent individuals and explore more valuable solution space regions, and the positions of the high-temperature water vapor individuals are updated by using the following formula: wherein, is the position of the i-th individual, is the position of the i+1-th individual, represents a random number between [0, 1], represents the position difference between the current elite individual and other worse individuals, is the current individual, is one of the individuals in the middle position of the entire population fitness value.

5. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 4, characterized in that: The exploration process updates the position of each individual and is represented as: wherein, the updated position of the individual, denotes the position of the current individual, denotes a random number vector based on Brownian motion, is a random number, is the current best solution, wherein and denote the second and third best individuals in the current population, respectively; is a randomly selected individual from the elite set, is the center position of the population; denotes the centroid position of the individuals whose fitness values are in the top 50%, and N1 is the number of leaders, i.e., half of the swarm.

6. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 5, characterized in that: The snowmelt coefficient is calculated according to the following formula: where t is the number of the current iteration, is the maximum value; The snowmelt rate is calculated according to the following formula: In the formula, T is the daily average temperature.

7. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker as claimed in claim 6, characterized in that: The development process updates the position of each individual and is represented as: wherein is a random number.

8. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 7, characterized in that: The execution condensation strategy updates the position of each individual and comprises: wherein, is the position of the random individual responsible for the tth iteration in the population, e increases with the number of iterations, is a random number between [−1, 1].

9. The SAO-BiLSTM-based load forecasting method for the power system of an icebreaker according to claim 1, characterized in that: The principal components with cumulative contribution rates greater than a set value are selected from the feature sequence, the environmental factor data is projected onto the selected principal components to obtain the reduced data, and the dynamic sparse kernel principal component analysis method is specifically used to achieve the above, and the method comprises the following steps: Based on the collected environmental factor data of the icebreaker to be subjected to load prediction, the real-time working condition intensity index is calculated: where IceThick(t) is the sea ice thickness, Wind(t) is the 10 meter wind speed, dv / dt is the icebreaker acceleration, , , are the corresponding weights. According to the real-time working condition intensity index, the feature sparsity is adaptively adjusted: wherein, is the sparsity size, is the minimum sparsity, is the maximum sparsity, k is the adjustment slope; A target function with the dimension reduction projection matrix W as an optimization variable is constructed, and the automatic selection of the principal components is realized through sparse constraint: wherein the regularized intensity dynamically adjusted with sparsity, is a kernel mapping function, is a feature sequence; A proximal gradient descent method is used to solve the target function to obtain the optimal dynamic sparse projection matrix W(t); The feature sequence is projected to the optimal dynamic sparse projection matrix W(t) to obtain the reduced data Z(t): 。

Images