A power load prediction system and method for a multi-energy electric propulsion vessel

Abstract

The application discloses a power load prediction system of a multi-energy electric propulsion ship, which comprises a load information acquisition module, a climate information acquisition module, an energy management global optimization strategy formulation module, a ship control center display device, a wireless local area network module and a big data cloud platform; the load information acquisition module and the climate information acquisition module transmit historical load data and marine climate data to the big data cloud platform; the energy management global optimization strategy formulation module is used for formulating an hourly sailing task in a future sailing process of the multi-energy electric propulsion ship and an average sailing speed under the sailing task; the big data cloud platform is used for receiving data information uploaded through the wireless local area network, completing construction of a training sample set and regression modeling of a prediction model, and finally outputting a power load prediction value at a to-be-predicted moment. The application constructs a prediction model through a rich training set and an improved algorithm, increases the accuracy of model prediction, and makes power load prediction more accurate.

Description

Technical Field

[0001] This invention relates to the field of multi-energy ship load prediction technology, specifically, to a power load prediction system and method for multi-energy electric propulsion ships. Background Technology

[0002] Current research on power load forecasting for multi-energy electric propulsion vessels is limited. Traditional power load forecasting methods cannot effectively address the impact of energy management strategies on these vessels, which involve varying hourly voyages, resulting in non-fixed speeds. Furthermore, the influence of ocean currents, wind speed, direction, temperature, and relative humidity on load forecasting is a significant concern. The limited time available for multi-energy electric propulsion vessels restricts the number of training samples for power load forecasting. Additionally, the numerous factors mentioned above exert complex influences on power consumption. When multi-energy electric propulsion vessels operate across sea areas, changing marine climate conditions lead to continuous variations in power load, further complicating power load forecasting and reducing its accuracy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a power load prediction system and method for multi-energy electric propulsion ships, thereby solving the technical problem that changes in marine climate conditions make it difficult to predict the constantly changing power load of ships.

[0004] This invention provides a power load forecasting system for multi-energy electric propulsion ships, comprising: a load information acquisition module, a climate information acquisition module, an energy management global optimization strategy formulation module, a ship control center display device, a wireless local area network module, and a big data cloud platform;

[0005] The load information acquisition module and the climate information acquisition module transmit historical load data and marine climate data of the sea areas traversed by the multi-energy electric propulsion ship to the big data cloud platform through the wireless local area network module.

[0006] The energy management global optimization strategy formulation module is connected to the big data cloud platform through the wireless local area network module, and is used to formulate hourly navigation tasks and average speeds under the navigation tasks for multi-energy electric propulsion ships during future navigation.

[0007] The big data cloud platform is connected to the wireless local area network module and is used to receive data information uploaded through the wireless local area network, complete the construction of the training sample set and the regression modeling of the prediction model, and finally output the predicted value of the power load at the time to be predicted.

[0008] Furthermore, the big data cloud platform includes: a data storage and cloud database integration module, a training sample set construction module, a data processing module, and a load prediction model establishment module.

[0009] The data storage and cloud database integration module stores the received ship data and uploads it to the cloud database.

[0010] The training sample set construction module selects the received ship data and chooses the average value of the parameters affecting the power load prediction results at each hour and the average speed at each hour determined by the global energy management optimization strategy as feature vectors to construct the training sample set.

[0011] The data processing module uses Euclidean distance as the evaluation criterion and adds the average hourly power load value under similar navigation scenarios to the feature vector to complete the dimensionality expansion of the feature vector of the entire training sample set.

[0012] The load prediction model building module constructs a load prediction model using a training sample set with expanded dimensionality.

[0013] This invention also provides a prediction method for a power load prediction system of a multi-energy electric propulsion ship, comprising the following steps:

[0014] Step 1: Obtain the average speed of the ship for each hour of its current total voyage, with the aim of minimizing costs;

[0015] Step 2: Construct a training sample dataset using historical ship navigation data. The dataset includes:

[0016] Hourly power load values ​​that affect the power load forecast results, hourly climate information data that affect the power load forecast results, and hourly average speed;

[0017] Step 3: Expand the dimension of each feature vector in the dataset. Specifically, obtain the output values ​​of several feature vectors similar to the current feature vector through Euclidean distance, and add these output values ​​to the current feature vector to expand the dimension of the current feature vector.

[0018] Step 4: Preprocess the expanded dataset;

[0019] Step 5: Perform cluster analysis on the preprocessed dataset to obtain datasets of several classes;

[0020] Step 6: Train the model for each of the several classes of datasets using the improved least squares support vector machine algorithm to obtain several load prediction models. The improved least squares support vector machine algorithm is an improvement on the least squares support vector machine algorithm by using the butterfly algorithm with dynamic switching probability.

[0021] Step 7: Construct a feature vector from the climate information data of the next moment, the average speed of the next moment, the power load value of the current moment, the power load value of the previous moment, and the power load value of the same moment in the previous day that corresponds to the next moment. Then, expand the dimension of the feature vector and preprocess it.

[0022] Step 8: Select the load prediction model corresponding to the training set with the highest similarity to the feature vector obtained in Step 7, and input the feature vector into the obtained load prediction model to predict the power load value at the next moment.

[0023] Furthermore, in step 6, the regularization parameter C, the standardization parameter σ, and the weight coefficient δ of the variant chi-square kernel function in the least squares support vector machine algorithm are optimized using the butterfly algorithm with dynamic switching probability.

[0024] Furthermore, the formula for calculating the dynamic switching probability in the butterfly algorithm with dynamic switching probability is as follows:

[0025]

[0026] Where, p d For dynamically changing switching probabilities, p max is the maximum switching probability, b is a constant coefficient, Curcount is the current iteration number, and Loopcount is the total number of algorithm iterations.

[0027] Furthermore, the optimization process of the butterfly algorithm with dynamic switching probability includes: using the average particle size and fitness variance as indicators to judge whether the butterfly algorithm converges to the extreme point; when the average particle size and fitness variance are both less than the set value, the butterfly algorithm has converged to the extreme point, the population position in the butterfly algorithm is reinitialized using the optimal point set strategy, and the dynamic switching probability is set to zero, so that each butterfly in the butterfly algorithm only performs local optimization.

[0028] Furthermore, the average particle size is set to 0.01, and the fitness variance is set to 0.001.

[0029] The beneficial effects of this invention are:

[0030] (1) Before a multi-energy electric propulsion vessel performs a navigation mission, the average speed of the vessel during the hourly navigation mission is determined by using a global optimization strategy of energy management. This navigation speed is then used as an important factor affecting the prediction of the power load of the multi-energy electric propulsion vessel at a certain time on a certain day in the future, so as to improve the accuracy of load prediction.

[0031] (2) The average power consumption of ships under similar navigation scenarios is highly referential. Therefore, in addition to the traditional method of using load values ​​under continuous and periodic time as the feature vector of the training set, this invention adds the average power consumption of ships under similar navigation conditions when passing through similar sea areas with similar wind speed, wind direction, ocean current speed, and ocean current direction to the feature vector, which can obtain more accurate load prediction results.

[0032] (3) This invention uses LS-SVM with a fusion of variant chi-square kernel function as the main prediction algorithm. Based on the traditional use of RBF kernel function as the kernel function of LS-SVM algorithm, the variant chi-square kernel function and RBF kernel function are fused together. While considering the absolute distance between samples, the relative distance between samples is also taken into consideration, thereby improving the prediction performance of the model.

[0033] (4) The basic butterfly optimization algorithm has a fixed switching probability, resulting in a low probability of local optimization patterns appearing in the later stages of the algorithm, making it impossible to effectively optimize around the extreme points. This invention dynamically changes the switching probability, with the switching probability value decreasing as the number of iterations increases. This allows each butterfly to quickly fly to the vicinity of the global optimum in the early stages of the algorithm, and to conduct local optimization around the global optimum in the later stages. After re-initializing the positions of each butterfly using the optimal point set strategy and forcing the algorithm to perform local optimization, the computational power wasted in the later stages of the algorithm due to convergence to the extreme points can be transformed into a process of traversing the entire solution space, thereby better finding the optimal values ​​of the parameters to be determined in the LS-SVM algorithm with the fusion variant chi-square kernel function. Attached Figure Description

[0034] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0035] Figure 1 This is a schematic diagram of the system structure according to a specific embodiment of the present invention;

[0036] Figure 2 This is a flowchart of the improved butterfly optimization algorithm according to a specific embodiment of the present invention;

[0037] Figure 3 This is an overall flowchart of a specific embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Modifications to the present invention in various equivalent forms all fall within the scope defined by the appended claims.

[0040] Taking a multi-energy electric propulsion ocean-going transport vessel as an example, the average power load consumption of the vessel during a cross-sea mission in a certain hour on a future day is predicted. The specific implementation method is as follows:

[0041] A power load prediction system for multi-energy electric propulsion ships, such as Figure 1 As shown, it includes: load information acquisition module 1, climate information acquisition module 2, energy management global optimization strategy formulation module 3, ship control center display equipment 4, wireless local area network module 5, and big data cloud platform 6;

[0042] Load information acquisition module 1 and climate information acquisition module 2 collect historical load data and marine climate data of the multi-energy electric propulsion ocean transport ship during operation and the sea areas it passes through. After collecting these data, they are transmitted to the big data cloud platform 6 via wireless local area network module 5. The sampled data is also directly transmitted to the ship control center display device 4 for technical personnel to refer to.

[0043] The energy management global optimization strategy formulation module 3 collects information on the diesel generator set, photovoltaic power generation system, composite energy storage device, electric propulsion load, and other marine loads of the multi-energy electric propulsion transport ship to formulate hourly navigation tasks and average speeds under these tasks during future voyages. The module 3 transmits the hourly navigation tasks to the ship control center display device 4 for technical personnel reference, and transmits the average speed under the hourly navigation tasks to the big data cloud platform 6 via the wireless local area network module 5, as an important factor influencing the power load forecast results.

[0044] The big data cloud platform 6 includes a data storage and cloud database integration module 7, a training sample set construction module 8, a data processing module 9, and a load prediction model establishment module 10. The data storage and cloud database integration module 7 stores the transport ship's load data and climate information data from various sea areas during the voyage, which are transmitted to the cloud platform by the wireless local area network module 5, and uploads them to the cloud database to achieve the purpose of data information sharing.

[0045] The training sample set construction module 8 will select the factors that affect the final prediction result of the power load forecast by selecting the received load data and climate information data as the sample input vector, and the output value is the power load forecast value. Then, according to the energy management global optimization strategy formulation module 3, the average speed of the transport ship under the hourly navigation task in the future navigation process is formulated as the feature vector to construct the training sample set.

[0046] Data processing module 9 uses Euclidean distance as the evaluation criterion. It incorporates the average power consumption of ships navigating similar sea areas with similar wind speeds, directions, ocean currents, and directions into the training feature vector and the current prediction time input vector, thus expanding the dimension of the feature vector and improving the accuracy of load prediction. It also performs data normalization and smoothing on the training sample set and the current prediction time input vector, reducing the computational load of the subsequent load prediction model building module 10 and improving the processing speed. Finally, it uses the K-MEANS algorithm for cluster analysis on the training sample set.

[0047] The load forecasting model building module 10 receives several typical sample sets obtained after normalization, smoothing and clustering by the data processing module. It performs regression modeling on these typical sample sets respectively through the improved least squares support vector machine algorithm. It also receives the input vector information of the current forecast time after processing by the data processing module, selects the training sample set with the highest similarity to train the forecasting model obtained after regression, and performs power load forecasting. The output power load forecast value is transmitted to the ship control center display device 4 through the wireless local area network module 5, so that the ship energy management system can adjust the specific output of each power supply end in the next hour according to the power load forecast value.

[0048] The specific forecasting method for the power load of multi-energy electric propulsion ocean-going transport ships includes the following steps:

[0049] Step 1: Before the multi-energy electric propulsion ocean-going transport vessel performs its mission across sea areas, the energy management global optimization strategy formulation module formulates a specific implementation plan for the average speed per hour during the future voyage, including the following steps:

[0050] Step 101: Establish mathematical models for each power supply end:

[0051] (1) Marine diesel generator set

[0052] The mathematical model for marine diesel generators is as follows:

[0053] Fuel = η FC ·P DG

[0054] In the formula, P DG η is the output power of the diesel generator. FC is the fuel consumption rate of the diesel generator under different load rates, in g / (kW·h); Fuel is the fuel consumed by the diesel generator during operation, in L.

[0055] (2) Photovoltaic power generation system

[0056] The model of the photovoltaic system used in this multi-energy electric propulsion ship is as follows:

[0057] P PV =η PV ·S PV ·Ir

[0058] In the formula, Ir is the solar radiation intensity; S PV η represents the area of ​​the solar panels in the photovoltaic system. PV This is a correction value for the photoelectric conversion efficiency based on the nominal operating temperature (NOCT) correction method for solar cells, which is related to the performance parameters, quality factor, and external environmental conditions of the photovoltaic power generation system.

[0059] (3) Composite energy storage device

[0060] The composite energy storage device used on this ocean-going transport vessel consists of battery packs and supercapacitors, and their models are as follows:

[0061]

[0062] In the formula, E B (t+1) and E B (t) represents the current capacity of the battery pack at times t+1 and t, respectively; η B，charge and η B，discharge These represent the charging and discharging efficiencies of the battery pack, respectively; P B (t) represents the output power of the battery pack at time t, where discharge is a positive value and charging is a negative value; Δt is the time interval.

[0063]

[0064] In the formula, E C (t+1) and E C (t) represents the current capacitance of the supercapacitor at times t+1 and t, respectively; ηC，charge and η C，dischange These represent the charging and discharging efficiencies of the supercapacitor, respectively; P C (t) represents the output power of the supercapacitor at time t, where discharge is a positive value and charging is a negative value.

[0065] Step 102: Analyze the constraints of the global energy management optimization strategy:

[0066] (1) Power balance constraint

[0067] P DG +P PV +P B +P C +P S =P M +P L

[0068] In the formula, P DG P is the output power of the marine diesel generator set. PV P represents the output power of the photovoltaic power generation system. B P is the output power of the battery pack. G P represents the output power of the supercapacitor. S P represents the output power of the shore power system. M The power requirements for an ideal propulsion motor; P L To meet the power requirements of other marine loads.

[0069] (2) Output power constraint of distributed power source

[0070]

[0071] In the formula, P i The output power of each distributed power source; These represent the upper and lower power limits for each distributed power source.

[0072] (3) Energy storage device capacity constraints

[0073]

[0074] In the formula, E i This represents the current capacity of each energy storage device; These represent the upper and lower limits of the capacity of energy storage devices, respectively.

[0075] (4) Range constraints

[0076]

[0077] In the formula, V S (t) represents the ship's real-time speed; Δt represents the time interval; DIS iLet be the expected planned voyage distance for the i-th segment of the voyage; ΔDIS is the allowable total voyage error.

[0078] Step 103: Determine the objective function optimized by the global optimization strategy for quantity management:

[0079] The objective function of the optimization model is expressed as follows:

[0080] F = f1(x) + f2(x)

[0081] f1 = C DG +C B +c C

[0082]

[0083] In the formula, F represents the total operating cost of the ship; f1(x) and f2(x) represent the economic cost and environmental cost of the ship, respectively, namely the operating cost of the electric power system of the multi-energy electric propulsion ship and the total pollution emissions of the ship; C DG C B C C , , respectively, represent the operating costs of the marine diesel generator set, battery pack, and supercapacitor; Emission is the total pollutant emissions from the marine diesel generator set, in grams.

[0084] c DG C B C C The specific representations of Emission are as follows:

[0085] C DG = price·Fuel + α(L) FH )·P DG

[0086] In the formula, price is the fuel price, $ / kg; Fuel is the real-time fuel consumption of the marine diesel generator; α is the operating and maintenance cost coefficient of the diesel generator, and L is the load factor. FH The function of P; DG This refers to the output power of the marine diesel generator.

[0087] C B =Deg B =β B ·(B CD +B E )

[0088] In the formula, Deg B Costs associated with battery pack degradation; β B B is the degradation loss factor of the battery pack; CD B represents the number of charge and discharge cycles of the battery pack.E This refers to the extent of excessive use of the battery pack.

[0089] C C =Deg C =β C ·C CD

[0090] In the formula, Deg C Cost of degradation losses for supercapacitors; β C C is the degradation loss factor of the supercapacitor; CD This refers to the number of charge and discharge cycles of the battery pack.

[0091]

[0092] In the formula, Fuel represents the real-time fuel consumption of the marine diesel generator; ξ i The conversion coefficient, in g / L, represents the atmospheric pollutants emitted by marine diesel generators when they consume fuel.

[0093] Step 104: By sampling the probability slices of the remaining power load usage to generate hourly load power consumption, this is added to the propulsion motor power required to perform the designated navigation task during different navigation periods to obtain the total power load demand. Then, the hourly navigation task and its required propulsion motor output power are optimized through a swarm intelligence algorithm, and the output of each power supply end is optimized under power constraints. Finally, based on the goal of minimizing ship operating costs, the optimal global optimization strategy for ship energy management is obtained.

[0094] Step 105: After obtaining the hourly output power of the propulsion motor required for each segment of the voyage in this scheme, the average speed of the ship during each hour of the total voyage can be obtained through the fitting formula shown below in the scheme with the lowest total operating cost in the global optimization strategy for energy management.

[0095]

[0096] In the formula, V s To determine the speed of a ship; P is a constant coefficient; M To increase the output power of the motor.

[0097] Step 2: Training Sample Set Construction Module. Considering that changes in marine climate such as wind speed and ocean currents can significantly impact the power consumption of ocean-going transport vessels when they perform missions across sea areas, and taking into account the similarity of hourly average power consumption in both continuous and periodic time, the input training features of the load are summarized. The 10-dimensional feature vector of the training sample is as follows: load value at the same time the day before the current prediction time on the prediction day, load value at two times before the current prediction time on the prediction day, load value at the time before the current prediction time on the prediction day, average temperature, average humidity, average relative wind speed, wind direction, average ocean current speed, ocean current direction, and average speed determined by the global energy management optimization strategy.

[0098] Step 3: The data processing module adds the load values ​​from similar navigation scenarios to the feature vector, thus expanding the dimension of the feature vector for the entire training sample set. The specific process is as follows:

[0099] Step 301: Let the number of training sample points be n, and select a sample point x from the training data sample set. i Let i = 1, 2, ..., n. Calculate the Euclidean distance between this sample point and all other sample points.

[0100] Step 302: Select the sample point x i Add the average load power consumption values ​​of the three closest sample points in Euclidean distance to sample point x. i In the eigenvectors.

[0101] Step 303: Repeat steps 301 and 302 until the dimension of the feature vector is expanded for each sample point in the sample set.

[0102] Suppose we expand the dimension of the feature vector of the sample point at 13:00 on June 23rd. After calculating the Euclidean distance, we find that the three sample points with the closest Euclidean distance to the sample point at 13:00 on June 23rd are the sample point at 9:00 on June 22nd, the sample point at 11:00 on June 17th, and the sample point at 12:00 on June 17th. Then, the feature vector of the sample point at 13:00 on June 23rd after dimension expansion is: the average power load consumption value at 12:00 on June 23rd, the average power load consumption value at 11:00 on June 23rd, the average power load consumption value at 13:00 on June 22nd, the average temperature, average humidity, average relative wind speed, wind direction, average ocean current speed, ocean current direction, the average speed determined by the global energy management optimization strategy at 13:00 on June 23rd, the average power load consumption value at 9:00 on June 22nd, the average power load consumption value at 11:00 on June 17th, and the average power load consumption value at 12:00 on June 17th.

[0103] Step 4: The data processing module smooths outliers in the training sample set load using the following formula:

[0104]

[0105]

[0106] Where K(d,t) represents the power load value at time t on day d, and δ(t) is the threshold.

[0107] The missing data is processed using the formula shown below:

[0108]

[0109] Where K(d,t) is the missing data at time t on day d, while K(d1,t) and K(d2,t) are the power load values ​​closest to day d and under the same operating conditions.

[0110] The data is normalized using the following formula:

[0111]

[0112] In the formula, x represents the data before normalization; x max x min y represents the maximum and minimum values ​​in the historical data, and y represents the normalized data.

[0113] Step 5: The data processing module uses the K-MEANS algorithm to perform cluster analysis on the training sample set, completing the data processing of the training sample set. Let the number of clusters after clustering be k, then in the total training sample set, we obtain k typical sample sets and k cluster centroids.

[0114] Step 6: The load forecasting model building module trains the clustered k typical sample sets using the improved least squares support vector machine algorithm, resulting in several different load forecasting models.

[0115] The specific implementation plan includes the following steps:

[0116] Step 601: Use the least squares support vector machine algorithm with a fusion variant of the chi-square kernel function as the prediction model. The specific process is as follows:

[0117] The basic principle of LS-SVM is to train sample data (x i y i ), i = 1, 2, ..., n. x i For input data, y i The output data is n, where n is the total number of samples. This is achieved through a non-linear mapping. The training samples are mapped to a high-dimensional feature space, and then linear regression is performed in the high-dimensional feature space. The regression function is then:

[0118]

[0119] In the formula: ω represents the weight vector, and b represents the regression bias. This represents a mapping function that maps samples from a low-dimensional space to a high-dimensional space.

[0120] Based on the principle of risk minimization, the regression constraints of LS-SVM are:

[0121]

[0122]

[0123] In the formula, e i ∈R l×1 Let R be the regression error, and e be the regression error. i The error vector consists of the errors, where C represents the regularization parameter, used to control the degree of penalty for errors during the regression process, ||ω||. 2 To control the complexity of the model, b is a constant and ω is a weight vector.

[0124] To facilitate solving the constraints, the Lagrangian function is introduced, transforming the constrained problem into an unconstrained problem:

[0125]

[0126] For solving the minimum problem within a specified scope with constraints, the KKT conditions can be obtained.

[0127]

[0128] Eliminating ω and e from the equation, the linear equation is as follows:

[0129]

[0130] In the formula, E = [1, 1, ..., 1] T , The kernel function represents the suitable conditions, I represents the identity matrix, and Y = [y1, y2, ..., y3]. l ] T Then the prediction model of LSSVM can be expressed as:

[0131]

[0132] Since most kernel functions are based on Euclidean distance, which reflects the absolute distance between samples but ignores the relative distance, while chi-square distance effectively reflects the relative distance, this invention will consider the impact of both distance metrics on the prediction results and fuse the variant chi-square kernel function and the RBF kernel function. The specific formula for the kernel function is as follows:

[0133]

[0134] In the formula, x i This represents the input vector, x j σ represents the training samples, σ represents the standardized parameters that determine the width of the center point around which the Gaussian function is surrounded, and δ is the weight coefficient of the fusion variant chi-square kernel function.

[0135] Step 602: Use the improved butterfly optimization algorithm to determine the regularization parameter C, standardization parameter σ, and weight coefficient δ in the LS-SVM algorithm with the fusion variant chi-square kernel function. The specific process is as follows:

[0136] First, the average relative error is selected as the fitness function for the butterfly optimization algorithm, and the specific formula is shown below:

[0137]

[0138] Where P(i) is the predicted load value, R(i) is the actual load value, and E MRE This represents the average relative error.

[0139] The mathematical formula for the global optimization mode of the butterfly optimization algorithm is shown below:

[0140]

[0141] In the formula, Let g represent the solution vector of the i-th butterfly in the t-th iteration, where r is a uniformly distributed random number between [0, 1]. * Let f represent the current global optimal solution, and f be the concentration of the butterfly's fragrance. The mathematical formula for f is as follows:

[0142] f = c * I a

[0143] Where f represents the concentration of the fragrance, I represents the intensity of the stimulus, which is related to the fitness; a represents the power exponent, and c represents the perceptual morphology. The range of a and c is usually [0, 1].

[0144] The mathematical formula for the local optimization mode of the butterfly optimization algorithm is shown below:

[0145]

[0146] In the formula, Let represent the solution vectors of the i-th, j-th, and k-th butterflies in the t-th generation, respectively. and It is the position vector of two individuals randomly selected in the t-th generation of the population.

[0147] The rand function is used to generate random numbers between 0 and 1. When the random number is less than the switching probability p (p is a set number of values ​​between 0 and 1), the butterfly optimization algorithm performs global optimization. When the random number is greater than or equal to the switching probability p, the butterfly optimization algorithm performs local optimization.

[0148] This invention utilizes a dynamic switching probability combined with the global and local optimization modes of the butterfly optimization algorithm, as shown in the following formula:

[0149]

[0150] In the formula P d For dynamically changing switching probabilities, p max is the maximum switching probability, b is a constant coefficient, Curcount is the current iteration number, and Loopcount is the total number of algorithm iterations.

[0151] In the later stages of the algorithm, the average particle size and fitness variance are introduced as indicators to judge whether the algorithm converges to the extreme point. The formula for the average particle size is shown below:

[0152]

[0153] In the formula, it is assumed that the current iteration number of the algorithm is k, n is the number of particles in the particle swarm; L is the value of the diagonal of the search space during the particle search process, and x ij This represents the coordinate value of the j-th dimension of the i-th particle; This represents the average value of all particles in dimension J.

[0154] The formula for fitness variance is as follows:

[0155]

[0156]

[0157] In the formula, f k i This represents the fitness value of the i-th particle in the population at the current iteration generation k. is the current average fitness value of the population; f represents the normalization scaling factor, used to limit the size of the variance.

[0158] Each of the average particle size and fitness variance is assigned a set value, denoted as D. setWith σ 2 set When the average particle size and fitness variance are less than the set value, i.e., D(k) < D set , σ 2 <σ 2 set At this point, the algorithm is considered to have converged to the extreme point, generally D set With σ 2 set Set to D set =0.01, σ 2 set =0.001. Then, the optimal point set strategy is used to reinitialize the positions of each butterfly, ensuring that the butterflies are evenly distributed in the search space. The specific method for selecting n optimal points in m-dimensional space is as follows:

[0159] First, construct a set of ideal points x' = (x1', x2', ..., xn) containing n points. j ',...,x n j = (1, 2, ..., n);

[0160] Subsequently, the i-th point x in the set of best points i '=(x i1 ', x i2 ',...,x ij ',…,x im '), j = (1, 2, ..., m). Each dimension x ij The value of ' is determined by the following formula:

[0161]

[0162] q is the smallest prime number that satisfies q > 2m + 3.

[0163] Finally, the following formula is used to ensure that each dimension of the optimal point set falls within the range of the function:

[0164] x ij '=Lb j +mod(x ij ',1)*(Ub j -Lb j )

[0165] Where, x j '∈[L,i=(1,2,…,n),Lb=(Lb1,…,Lb j ,...,Lbm), Ub=(Ub1,...,Ub j , ..., Ub m )).

[0166] Since the optimization targets are the regularization parameter C, the standardization parameter σ, and the weight coefficient δ, m is set to 3. Lb and Ub are the lower and upper limits of the three parameters, respectively, so Lb = (0.1, 0.1, 0) and Ub = (200, 20, 1). The population size is set to 50, so n = 50.

[0167] After reinitializing the population using the best-set strategy, force p... d Setting the value to 0 allows each butterfly to perform only local optimization. This improvement transforms the computational effort wasted in the later stages of the algorithm due to convergence to extreme points into a process of traversing the entire solution space. Please refer to the flowchart of the improved Butterfly Optimization Algorithm (IBOA). Figure 2 .

[0168] Step 603: Perform IBOA-optimized LSSVM regression modeling on the data in each of the k typical sample sets to obtain k different load prediction models.

[0169] Step 7: The load forecasting model building module receives the current forecast time input vector information after dimensionality expansion, normalization, and smoothing by the data processing module. It then selects the typical sample set with the highest similarity to train the regression-derived forecasting model for power load forecasting. The specific process is as follows:

[0170] Step 701: Calculate the Euclidean distance between the input vector at the current prediction time and the sample points in the training sample set before data processing. Select the average load power consumption values ​​at the times corresponding to the three sample points with the closest Euclidean distance and add them to the input vector at the current prediction time, so that the dimension of the input vector at the prediction time is equal to that of the feature vector of the training sample set.

[0171] Step 702: Calculate the Euclidean distance between the input vector at the current prediction time and the centroids of each of the k typical sample sets.

[0172] Step 703: Assuming the class center with the smallest Euclidean distance from the input vector at the current prediction time belongs to the typical sample set of class i, then use the prediction model of class i to predict the input vector at the current prediction time. Please refer to the complete flowchart of the power load forecasting method. Figure 3 .

[0173] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A power load prediction system for multi-energy electric propulsion ships, characterized in that, include: Load information acquisition module, climate information acquisition module, energy management global optimization strategy formulation module, ship control center display equipment, wireless local area network module and big data cloud platform; The load information acquisition module and the climate information acquisition module transmit historical load data and marine climate data of the sea areas traversed by the multi-energy electric propulsion ship to the big data cloud platform through the wireless local area network module. The energy management global optimization strategy formulation module is connected to the big data cloud platform through the wireless local area network module, and is used to formulate hourly navigation tasks and average speeds under the navigation tasks for multi-energy electric propulsion ships during future navigation. The big data cloud platform is connected to the wireless local area network module and is used to receive data information uploaded through the wireless local area network, complete the construction of the training sample set, perform cluster analysis on the training sample set to obtain several typical sample sets, and perform regression modeling of the prediction model on each typical sample set by using the least squares support vector machine algorithm that integrates the RBF kernel function and the variant chi-square kernel function. Based on the received data information, the prediction model obtained after training the regression of the training sample set with the highest similarity is selected to perform power load prediction. The formula for the fused kernel function is as follows: ； In the formula, This represents the input vector. This represents the training samples. This represents the standardized parameters that determine the width of the Gaussian function around its center point. To incorporate the weighting coefficients of the variant chi-square kernel function, This is a variant of the chi-square kernel function.

2. The power load prediction system for multi-energy electric propulsion ships as described in claim 1, characterized in that, The big data cloud platform includes: a data storage and cloud database integration module, a training sample set construction module, a data processing module, and a load prediction model establishment module; The data storage and cloud database integration module stores the received ship data and uploads it to the cloud database. The training sample set construction module selects the received ship data and chooses the parameters that affect the power load prediction results, such as the average temperature, average humidity, average relative wind speed, wind direction, average ocean current speed, ocean current direction, and the average hourly speed of the ship as determined by the global energy management optimization strategy, as feature vectors to construct the training sample set. The data processing module uses Euclidean distance as the evaluation criterion and adds the hourly average power load value under similar navigation scenarios to the feature vector to complete the dimensionality expansion of the feature vector of the entire training sample set; the load prediction model building module constructs a load prediction model through the dimensionally expanded training sample set.

3. A prediction method for a power load prediction system of a multi-energy electric propulsion ship, applicable to the power load prediction system of a multi-energy electric propulsion ship as described in claim 1 or 2, characterized in that, The prediction method of the power load prediction system for multi-energy electric propulsion ships includes the following steps: Step 1: With the goal of minimizing cost, obtain the average speed of the ship for each hour during its current total voyage based on the global energy management optimization strategy. The specific formula is as follows: ； In the formula, To determine the speed of a ship; The coefficients are constants. To increase the output power of the motor; Step 2: Construct a training sample dataset using historical ship navigation data. The dataset includes: hourly power load values ​​that affect the power load prediction results, hourly climate information data that affect the power load prediction results, and hourly average speed. Step 3: Expand the dimension of each feature vector in the dataset. Specifically, obtain the output values ​​of several feature vectors similar to the current feature vector through Euclidean distance, and add these output values ​​to the current feature vector to expand its dimension. Step 4: Preprocess the expanded dataset; Step 5: Perform cluster analysis on the preprocessed dataset to obtain datasets of several classes; Step 6: Train the model for each of the several classes of datasets using the improved least squares support vector machine algorithm to obtain several load prediction models. The improved least squares support vector machine algorithm is an improvement on the least squares support vector machine algorithm by using the butterfly algorithm with dynamic switching probability. Step 7: Construct a feature vector from the climate information data of the next moment, the average speed of the next moment, the power load value of the current moment, the power load value of the previous moment, and the power load value of the same moment in the previous day that corresponds to the next moment. Then, expand the dimension of the feature vector and preprocess it. Step 8: Select the load prediction model corresponding to the training set with the highest similarity to the feature vector obtained in Step 7, and input the feature vector into the obtained load prediction model to predict the power load value at the next moment.

4. The prediction method of the power load prediction system for multi-energy electric propulsion ships as described in claim 3, characterized in that, In step 6, the regularization parameter C, the standardization parameter, and the weight coefficients of the fusion variant chi-square kernel function in the least squares support vector machine algorithm are optimized using the butterfly algorithm with dynamic switching probability. The variant chi-square kernel function is a kernel function that uses chi-square distance as a distance metric and maps chi-square distance to similarity in the form of an exponential function.

5. The prediction method of the power load prediction system for multi-energy electric propulsion ships as described in claim 4, characterized in that, The formula for calculating the dynamic switching probability in the butterfly algorithm with dynamic switching probability is as follows: ； in, The switching probability changes dynamically. For the maximum switching probability, The constant coefficients, This represents the current iteration number. This represents the total number of algorithm iterations.

6. The prediction method of the power load prediction system for multi-energy electric propulsion ships as described in claim 5, characterized in that, The optimization process of the butterfly algorithm with dynamic switching probability includes: using the average particle size and fitness variance as indicators to judge whether the butterfly algorithm has converged to the extreme point; when the average particle size and fitness variance are both less than the set value, the butterfly algorithm has converged to the extreme point, the best point set strategy is used to reinitialize the population position in the butterfly algorithm, and the dynamic switching probability is set to zero, so that each butterfly in the butterfly algorithm only performs local optimization.

7. The prediction method of the power load prediction system for multi-energy electric propulsion ships as described in claim 6, characterized in that, The average particle size is set to 0.01; the fitness variance is set to 0.001.

Images