Energy efficiency evaluation and optimization method and system for photovoltaic power generation and cold storage system

Abstract

The present application provides an energy efficiency conversion assessment method for a photovoltaic power generation and cold storage system, comprising: obtaining real-time power data and weather condition data of a photovoltaic power station, and calculating a power fluctuation rate sequence; extracting characteristic values ​​from the power fluctuation rate sequence, and determining a sequence of influence coefficients of fluctuations on ice-making energy input; obtaining temperature data and water volume data in an ice storage tank, updating a real-time energy input model using the influence coefficient sequence, and calculating dynamic changes in the temperature sequence and the water volume sequence; constructing a cooling efficiency sequence based on the dynamic changes in the temperature sequence and the water volume sequence; obtaining building load demand data and ambient temperature data, and calculating a load demand change rate sequence; constructing a dynamic coupling relationship sequence using the cooling efficiency sequence and the load demand change rate sequence, and determining an energy transfer delay sequence; and extracting full-cycle features from the dynamic coupling relationship sequence and the energy transfer delay sequence to generate an energy efficiency prediction sequence.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an energy efficiency evaluation and optimization method and system for a photovoltaic power generation and cold storage system. Background Art

[0002] Existing methods for energy efficiency evaluation and optimization of photovoltaic (PV) thermal storage systems generally focus on static models, failing to fully consider the complex interplay between the volatility of PV output power and the dynamic operation of the thermal storage system. PV power, driven by external factors such as weather conditions and solar insolation, exhibits significant uncertainty and time-variability. This fluctuation directly impacts the energy input to the ice-making and thermal storage process, making it difficult to stabilize the real-time state of the water temperature and volume within the ice storage tank. Dynamic changes in water temperature and volume further impact the efficiency of the ice-melting and cooling process, which depends not only on the current state of the ice storage tank but is also closely coupled to the real-time demand of the external cooling load. Load demand can fluctuate dramatically over short periods of time due to factors such as building usage patterns and ambient temperature fluctuations, leading to time lags or mismatches in the energy transfer between ice-making and cooling. Traditional static models typically assume fixed values ​​for parameters such as input power, water temperature, and water volume, ignoring the nonlinear variations of these variables in actual operation and the feedback effects between them. This assumption results in evaluation results that fail to reflect the system's true energy efficiency performance over varying timescales, with significant deviations particularly noticeable under extreme weather conditions or sudden load changes. Therefore, it is necessary to accurately quantify the instantaneous impact of photovoltaic power generation volatility on the energy input of ice making and cold storage, while considering the dynamic coupling relationship between the water temperature and water volume of the ice storage tank and the cooling load demand. On this basis, the limitations of traditional static models in multi-variable real-time interaction scenarios should be overcome, and an evaluation framework that can capture the energy efficiency characteristics of the system throughout its life cycle should be constructed. Summary of the Invention

[0003] The present invention provides an energy efficiency evaluation and optimization method for a photovoltaic power generation cold storage system, which mainly includes:

[0004] Real-time power data and weather condition data of the photovoltaic power station are obtained to calculate a power fluctuation rate sequence; characteristic values ​​are extracted from the power fluctuation rate sequence to determine a sequence of influence coefficients of fluctuations on ice-making energy input; temperature data and water volume data in the ice storage tank are obtained, a real-time energy input model is updated using the influence coefficient sequence, and dynamic changes in the temperature sequence and the water volume sequence are calculated; a cooling efficiency sequence is constructed based on the dynamic changes in the temperature sequence and the water volume sequence; building load demand data and ambient temperature data are obtained to calculate a load demand change rate sequence; a dynamic coupling relationship sequence is constructed using the cooling efficiency sequence and the load demand change rate sequence to determine an energy transfer delay sequence; full-cycle features are extracted from the dynamic coupling relationship sequence and the energy transfer delay sequence to generate an energy efficiency prediction sequence.

[0005] Furthermore, the method of obtaining real-time power data and weather condition data of the photovoltaic power station and calculating the power fluctuation sequence includes: collecting photovoltaic output power and sunshine intensity through sensors, calculating the power fluctuation rate using a time series analysis method, and setting the power fluctuation rate to a preset value if the power at the previous time point is zero and the current power is non-zero; processing the power fluctuation rate and sunshine intensity through a sliding window method, calculating the average power fluctuation rate and the average sunshine intensity in a continuous time period, and analyzing the correlation between the average power fluctuation rate and the average sunshine intensity using the Pearson correlation coefficient to obtain a quantitative result of the power fluctuation.

[0006] Furthermore, the method of extracting characteristic values ​​from the power fluctuation rate sequence and determining a sequence of influence coefficients of fluctuations on ice-making energy input includes: processing the power fluctuation rate sequence through a sliding window method to extract a smooth fluctuation characteristic sequence; calculating the energy input per unit time through integration, and using a convolutional neural network to analyze the instantaneous mapping relationship between the smooth fluctuation characteristic sequence and the energy input to obtain a mapping characteristic matrix; calculating the influence coefficient sequence based on the mapping characteristic matrix; if the influence coefficient exceeds a preset threshold, processing the relationship between the smooth fluctuation characteristic sequence and the energy input through a random forest algorithm, adjusting the influence coefficient sequence, and obtaining a dynamic influence sequence of fluctuations on energy input.

[0007] Furthermore, the real-time energy input model is updated through the influence coefficient sequence to calculate the dynamic changes of the temperature sequence and the water volume sequence, including: updating the energy sequence through the influence coefficient sequence and the energy input calculation, and generating the temperature sequence and the water volume sequence by using the differential equation; processing the temperature sequence and the water volume sequence by the sliding window method to obtain the smoothed temperature sequence and the smoothed water volume sequence; using linear regression to analyze the mapping relationship between the smoothed temperature sequence and the updated energy sequence to obtain the temperature-energy correlation coefficient matrix; combining the temperature-energy correlation coefficient matrix and the water volume energy correlation coefficient matrix by the matrix fusion method to generate a comprehensive dynamic influence matrix, and adjusting the temperature sequence and the water volume sequence according to the comprehensive dynamic influence matrix.

[0008] Furthermore, the cooling efficiency sequence is constructed according to the dynamic changes of the temperature sequence and the water quantity sequence, including: calculating the initial cooling efficiency sequence through a preset efficiency weight coefficient, inputting the initial cooling efficiency sequence and historical data into a support vector machine model to generate an efficiency prediction sequence; the efficiency weight coefficient is determined by multivariate linear regression of historical cooling data, wherein the temperature weight γ has a value range of 0.32-0.45, and the water quantity weight has a value range of =0.55-0.68; processing the efficiency prediction sequence through a sliding window method to obtain a smoothed efficiency sequence; using the least squares method to analyze the linear relationship between the smoothed efficiency sequence and the temperature sequence to obtain a temperature efficiency correlation coefficient matrix; generating a comprehensive efficiency impact matrix through matrix weighted fusion of the temperature efficiency correlation coefficient matrix and the water quantity efficiency correlation coefficient matrix, and adjusting the smoothed efficiency sequence according to the comprehensive efficiency impact matrix.

[0009] Furthermore, the method of obtaining building load demand data and ambient temperature data and calculating the load demand change rate sequence includes: decomposing the load demand data through a long short-term memory network to obtain trend components and periodic components, and calculating the load demand change rate sequence; using the Pearson correlation coefficient to analyze the correlation between the load demand change rate sequence and the ambient temperature data to obtain a correlation coefficient sequence; processing the correlation coefficient sequence through a sliding window method to obtain a smoothed correlation coefficient sequence; using a random forest model to fit the nonlinear relationship between the smoothed correlation coefficient sequence and the ambient temperature data to generate a fitting relationship function, and adjusting the load demand change rate sequence according to the fitting relationship function.

[0010] Furthermore, the cooling efficiency sequence and the load demand change rate sequence are used to construct a dynamic coupling relationship sequence and determine an energy transfer delay sequence, including: calculating the dynamic coupling relationship sequence through a preset coupling weight; if the dynamic coupling relationship sequence exceeds a preset threshold, calculating the time lag effect characteristics through a sliding window method to generate an energy transfer delay sequence; using a support vector machine algorithm to process the dynamic coupling relationship sequence and the energy transfer delay sequence to obtain a mapping relationship function; analyzing the mapping relationship function and the cooling efficiency sequence through a long short-term memory network, predicting the dynamic coupling relationship sequence of the next time period, and adjusting the dynamic coupling relationship sequence according to the prediction result.

[0011] Furthermore, the full-cycle features are extracted from the dynamic coupling relationship sequence and the energy transfer delay sequence to generate an energy efficiency prediction sequence, including: extracting the periodic features of the dynamic coupling relationship sequence and the energy transfer delay sequence through a preprocessing method, using a random forest algorithm to perform multivariate regression on the temperature sequence, water volume sequence, cooling efficiency sequence and load demand change rate sequence to generate a preliminary energy efficiency prediction sequence; calculating the fluctuation feature sequence of the preliminary energy efficiency prediction sequence through a sliding window method, if the fluctuation feature sequence exceeds a preset threshold, using a support vector machine algorithm to process the fluctuation feature sequence and the energy transfer delay sequence to obtain a mapping feature sequence, and adjusting the energy efficiency prediction sequence according to the mapping feature sequence.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] This invention discloses a method for evaluating and optimizing the energy efficiency of a photovoltaic power generation and cold storage system. This method collects real-time photovoltaic power and weather data, analyzes the impact of power fluctuations on ice-making energy input, and constructs a dynamic ice-melting cooling efficiency model based on changes in ice storage tank temperature and water volume. Furthermore, the system uses deep learning methods to predict building load demand, establishes a dynamic coupling relationship between cooling efficiency and load demand, and accounts for time lag effects. The invention also employs multivariate regression analysis and adaptive filtering techniques to achieve energy efficiency prediction and model parameter optimization at different time scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention provides a flow chart of an energy efficiency evaluation and optimization method for a photovoltaic power generation and cold storage system. DETAILED DESCRIPTION

[0015] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0016] like Figure 1 In this embodiment, an energy efficiency evaluation and optimization method for a photovoltaic power generation and cold storage system may specifically include:

[0017] S101. Obtain real-time power data and weather condition data of the photovoltaic power station. Use sensors to collect the photovoltaic output power P(t) and sunshine intensity I(t) every minute. Use time series analysis to calculate the power fluctuation rate F(t) = |P(t) - P(t-1)| / P(t-1) to obtain the quantitative results of power fluctuation.

[0018] Sensors collect photovoltaic output power P(t) and sunshine intensity I(t) every minute, where P(t) is the power at the current time and P(t-1) is the power from the previous minute. The power fluctuation F(t) is calculated as |P(t)-P(t-1)| / P(t-1). If P(t-1) is 0 and P(t) is non-zero, F(t) is set to 0. The sunshine intensity I(t) in the time series is obtained, and the correlation coefficient between the power fluctuation F(t) and the sunshine intensity I(t) is calculated using the Pearson correlation coefficient. If the power fluctuation F(t) exceeds a preset threshold based on statistical analysis of historical data, the fluctuation is considered abnormal. The time series is processed using a sliding window method with a window size of 5 minutes and a step size of 1 minute. The average power fluctuation and average sunshine intensity over 5 consecutive minutes are calculated to obtain the smoothed fluctuation trend and average sunshine intensity. The correlation coefficient between the smoothed fluctuation trend and average sunshine intensity is calculated using the Pearson correlation coefficient to determine the relationship between the fluctuation and weather conditions.

[0019] For example, by collecting photovoltaic output power and sunshine intensity data every minute, sensors can monitor and analyze the operating status of the photovoltaic system in real time. Power fluctuation, as an important indicator of photovoltaic output power changes, can reflect the stability of system operation, while sunshine intensity is one of the main external factors affecting power fluctuations. The following detailed analysis and examples illustrate technical topics such as power fluctuation calculation, Pearson correlation coefficient analysis, abnormal fluctuation determination, and sliding window smoothing. As an implementation method, calculating the power fluctuation rate F(t) can intuitively reflect the dynamic changes in photovoltaic power. For example, suppose that on a clear morning, the power P(t-1) is 100 kilowatts in one minute and 110 kilowatts in the next minute. Then F(t) = |110-100| / 100 = 0.1, indicating a power fluctuation rate of 10%. If P(t-1) is 0 kilowatts in a certain minute and P(t) is 50 kilowatts, then F(t) is set to 0 by definition. This calculation method effectively captures sudden changes in power from zero to non-zero, such as when a PV system is first started up at sunrise. This method provides basic data for subsequent fluctuation analysis. Furthermore, using the Pearson correlation coefficient to calculate the correlation between power fluctuation F(t) and sunshine intensity I(t) reveals the degree of correlation between the two. For example, suppose 60 minutes of data were collected on a given day. During a certain period, sunshine intensity I(t) gradually increased from 500 watts / square meter to 800 watts / square meter, and power fluctuation F(t) also exhibited a certain upward trend. The Pearson correlation coefficient between the two may be close to 0.7, indicating that changes in sunshine intensity have a strong positive impact on power fluctuations. This analysis helps understand the impact of weather conditions on PV system operational stability and provides a basis for optimizing control strategies. Specifically, for identifying abnormal fluctuations, preset thresholds can be obtained through statistical analysis of historical data. Assuming that, through extensive historical data analysis, the power fluctuation rate F(t) normally has an average value of 0.15 and a standard deviation of 0.05, the threshold can be set to 0.25, which is the mean plus two standard deviations. If the calculated value of F(t) for a particular minute is 0.3, exceeding the threshold, it is considered an abnormal fluctuation. For example, in cloudy weather, the rapid movement of clouds causes a sharp change in sunlight intensity, causing the power to drop sharply from 200 kilowatts to 50 kilowatts, and F(t) to reach 0.75, which is clearly abnormal. This judgment mechanism can promptly detect potential problems, such as equipment failure or sudden weather changes, and provide early warnings to operations and maintenance personnel. It should be noted that the introduction of the sliding window method can further smooth the data, reduce noise interference, and thus more clearly observe the fluctuation trend. For example, assuming the window size is 5 minutes and the step size is 1 minute, the power fluctuation rates F(t) within a certain 5-minute period are 0.1, 0.12, 0.15, 0.09, and 0.11, respectively. The average fluctuation rate within the window is (0.1+0.12+0.15+0.09+0.11) / 5=0.114.At the same time, the average sunshine intensity within the window is calculated. For example, the corresponding values ​​are 600, 610, 590, 605, and 595 watts per square meter, with an average of 600 watts per square meter. This smoothing process produces a more stable fluctuation trend and sunshine intensity curve. Furthermore, recalculating the Pearson correlation coefficient on the smoothed data more accurately reflects the relationship between the fluctuation trend and sunshine intensity. For example, if the smoothed fluctuation trend shows a strong negative correlation with the average sunshine intensity, the correlation coefficient may be -0.6, indicating that the more stable the sunshine intensity, the smaller the power fluctuation. This analysis result can be used to optimize the power prediction model of the photovoltaic system and improve prediction accuracy. For example, in actual application, a photovoltaic power station used the above method to find that during the period of frequent summer afternoon thunderstorms, the correlation between the smoothed fluctuation trend and sunshine intensity was low, and the fluctuation rate frequently exceeded the threshold. After further analysis combined with meteorological data, the operations and maintenance team optimized the inverter response strategy to better adapt to short-term weather changes and reduce the number of abnormal downtimes.

[0020] S102. Extract characteristic values ​​from the power fluctuation rate F(t), calculate the energy input per unit time E(t) = ∫P(t)dt based on the ice-making equipment operation log, use a convolutional neural network to process the instantaneous mapping relationship between F(t) and E(t), and determine the influence coefficient K(t) of the fluctuation on the ice-making energy input.

[0021] A sliding window, F(t), with a window size of 3 minutes and a step size of 1 minute, is used to process the data, resulting in a smoothed fluctuation sequence. Eigenvalues ​​are extracted from the smoothed fluctuation sequence to form a fluctuation feature sequence. The energy per unit time, E(t) = ∫P(t)dt, is calculated by integration. A convolutional neural network is used to analyze the instantaneous mapping relationship between the fluctuation feature sequence and E(t). This yields a mapping feature matrix containing the convolution kernel weights and the output of the pooling layer. The convolution kernel has a size of 3×1, a step size of 1, and a Reinforced Luminance (ReLU) activation function. The influence coefficient K(t) is calculated for the mapping feature matrix. If K(t) exceeds a preset threshold of 2, the fluctuation is deemed to have a significant impact on the energy input, and the mapping feature matrix is ​​updated. The relationship between the smoothed fluctuation sequence and the energy per unit time, E(t), is processed using a random forest algorithm to obtain an energy fluctuation distribution. The influence coefficient K(t) is adjusted based on the energy fluctuation distribution. If the deviation between the adjusted K(t) and the instantaneous mapping relationship is less than a preset threshold of 1, the dynamic impact of the fluctuation on the ice-making energy input is determined. The adjusted K(t) is jointly analyzed with the time series data through a convolutional neural network to obtain the final volatility influence coefficient sequence.

[0022] As an example, as an implementation, using a sliding window process F(t) can smooth data and reduce noise interference. Assuming a window size of 3 minutes and a step size of 1 minute, and the power fluctuations within a 3-minute period are 0.1, 0.12, and 0.15, respectively, the average fluctuation within the window is 0.123. This smoothing process can produce a more stable fluctuation trend, facilitating subsequent analysis. Furthermore, extracting feature values ​​from the smoothed fluctuation sequence to form a fluctuation feature sequence can capture key fluctuation patterns. For example, the maximum, minimum, and mean values ​​within each window can be extracted as features, which can reflect the intensity and distribution of the fluctuations. By integrating and calculating the energy per unit time E(t), the actual output performance of the photovoltaic system can be evaluated. Assuming that the power curve approximates a parabola within a 5-minute period, the total energy output during this period can be obtained through integration. It should be noted that using a convolutional neural network to analyze the instantaneous mapping between the fluctuation feature sequence and E(t) can capture complex nonlinear relationships. The network may contain multiple convolutional and pooling layers to extract local and global features of the time series. Specifically, the convolutional neural network comprises three convolutional layers with 16 / 32 / 64 filters, respectively. The pooling layer employs max pooling with a pooling size of 2×1. The resulting mapping feature matrix reflects the pattern of fluctuations affecting energy output. As an implementation, an influence coefficient K(t) is calculated for the mapping feature matrix to quantify the extent of fluctuations' impact on energy input. If K(t) exceeds a preset threshold of 2, indicating a significant impact of fluctuations on energy input, the mapping feature matrix needs to be updated to adapt to the new relationship pattern. This dynamic adjustment mechanism can improve the adaptability and accuracy of the model. Specifically, the relationship between the smoothed fluctuation sequence and the energy per unit time, E(t), is processed using a random forest algorithm to obtain an energy fluctuation distribution. The advantage of random forests is their ability to handle nonlinear relationships and multidimensional features, making them suitable for analyzing complex energy fluctuation patterns. Based on the resulting distribution, the influence coefficient K(t) can be further adjusted to better reflect the actual situation. Furthermore, the adjusted K(t) is combined with time series data using a convolutional neural network to obtain the final fluctuation influence coefficient sequence. This method comprehensively considers the characteristics of time series and the impact of fluctuations, and can more accurately describe the dynamic impact of fluctuations on ice-making energy input.

[0023] S103. Obtain the temperature sensor data T(t) and the water volume sensor data V(t) in the ice storage tank, update the real-time energy input model E'(t)=K(t)×E(t) through the influence coefficient K(t), and use the differential equations T'(t)=T(t-1)+α×E'(t) and V'(t)=V(t-1)+β×E'(t) to calculate the dynamic changes of water temperature and water volume, where α and β are preset heat transfer and water storage coefficients.

[0024] Temperature data T(t) and water volume data V(t) are collected from the ice storage tank once per minute. The updated energy E'(t) = K(t) × E(t) is calculated using a predefined influence coefficient K(t), where E(t) is the current energy value and K(t) is the influence coefficient at time t. The temperature series T'(t) and water volume series V'(t) are generated using the differential equations T'(t) = T(t-1) + α × E'(t) and V'(t) = V(t-1) + β × E'(t), respectively, where α and β are the predetermined temperature and water volume response coefficients to energy, respectively. The temperature series T'(t) and water volume series V'(t) are processed using a sliding window with a window size of 5 minutes and a step size of 1 minute to obtain smoothed temperature and water volume series. Linear regression is used to analyze the mapping between the smoothed temperature series and the updated energy E'(t). The temperature-energy correlation coefficient matrix is ​​then derived based on the time variation t. Linear regression is used to process the dynamic relationship between the smoothed water volume sequence and the updated energy E'(t). Combined with the time change t, a water volume energy correlation coefficient matrix is ​​obtained. The temperature energy correlation coefficient matrix and the water volume energy correlation coefficient matrix are fused through matrix multiplication and combined with the influence coefficient K(t) to obtain a comprehensive dynamic influence matrix. If the element value in the comprehensive dynamic influence matrix exceeds the preset threshold of 3, the temperature sequence T'(t) and the water volume sequence V'(t) are updated using a differential equation to obtain an adjusted dynamic change sequence. Based on the adjusted dynamic change sequence, the temperature sequence T'(t), the water volume sequence V'(t), and the updated energy E'(t) are fused using a weighted average method. The weights are determined based on the predetermined contribution ratios of temperature and water volume to energy, resulting in a warm water energy distribution sequence within the ice storage tank.

[0025] For example, the collection and processing of temperature and water volume data within the ice storage tank, adjusting energy input using preset influence coefficients, can more accurately reflect the actual operating status of the system. As one embodiment, temperature T(t) and water volume V(t) data can be acquired at a frequency of once per minute. This high-frequency acquisition helps capture transient changes in the system. Specifically, energy E(t) is corrected using the influence coefficient K(t) to obtain an updated energy E'(t). This correction mechanism dynamically adjusts energy input to meet system demands at different times. For example, during peak ice-making periods, K(t) may be higher, resulting in an increase in E'(t), thereby improving ice-making efficiency. Furthermore, using difference equations to generate temperature series T'(t) and water volume series V'(t) can simulate the dynamic response of the system. The energy response coefficients α and β of temperature and water volume reflect the physical characteristics of the system. As one embodiment, these coefficients can be determined experimentally. For example, α might be 0.02°C / kWh and β might be 0.05L / kWh. It should be noted that using sliding window processing can smooth data fluctuations and improve data reliability. The window size is set to 5 minutes, with a step size of 1 minute. This setting preserves sufficient detailed information while effectively reducing noise interference. Linear regression analysis is used to explore the relationship between temperature, water volume, and energy. This method can quantify the mutual influence of various parameters. For example, it may be found that for every 1°C increase in temperature, energy consumption increases by 2 kWh. By integrating the effects of temperature and water volume through matrix operations, a comprehensive dynamic influence matrix can be obtained. As an implementation, a threshold of 3 can be set. When the matrix element value exceeds this threshold, it indicates that the system state has changed significantly and requires timely adjustment. This mechanism can improve the system's responsiveness and adaptability. Finally, the temperature, water volume, and energy data are integrated using a weighted average method to obtain the warm water energy distribution series within the ice storage tank. This comprehensive approach, which considers multiple factors, provides a more comprehensive description of the system state. For example, the weights for temperature, water volume, and energy can be set to 0.3, 0.3, and 0.4, respectively, to balance the influence of each factor. Through this data processing and analysis method, precise control and optimized operation of the ice-making system can be achieved, energy utilization efficiency can be improved, and operating costs can be reduced.

[0026] S104. Based on the dynamic change data of water temperature T'(t) and water volume V'(t), construct an ice melting cooling efficiency function η(t) = γ × T'(t) + δ × V'(t), where γ and δ are preset efficiency weight coefficients. The support vector machine model is trained using historical cooling data to obtain a predicted value of η(t).

[0027] The dynamic series of water temperature T'(t) and water volume V'(t) in the ice storage tank are obtained. Using the preset temperature efficiency weight coefficient γ and water volume weight coefficient δ, the efficiency function η(t) = γ × T'(t) + δ × V'(t) is calculated to obtain the initial cooling efficiency series η(t). This initial cooling efficiency series η(t) and historical data are input into a support vector machine model for training to obtain the efficiency prediction series η'(t). The efficiency prediction series η'(t) is processed using a sliding window with a window size of 10 minutes and a step size of 1 minute, and mean smoothing is used to obtain the smoothed efficiency series η''(t). The smoothed efficiency series η''(t) is linearly regressed with the water temperature series T'(t) using the least squares method to obtain the temperature efficiency correlation coefficient matrix A(t). The smoothed efficiency series η''(t) is linearly regressed with the water volume series V'(t) using the least squares method to obtain the water volume efficiency correlation coefficient matrix B(t). The temperature efficiency correlation coefficient matrix A(t) and the water efficiency correlation coefficient matrix B(t) are weighted and fused, with the weights determined by the preset temperature weight γ and water weight δ, to obtain the comprehensive efficiency impact matrix C(t). If the value of an element in the comprehensive efficiency impact matrix C(t) exceeds the preset threshold of 5, the smoothed efficiency sequence is updated using the difference equation η'''(t)=η''(t-1)+λ×C(t) to obtain the adjusted efficiency sequence η'''(t), where λ is the preset adjustment coefficient.

[0028] For example, by obtaining the dynamic series of water temperature T'(t) and water volume V'(t) in the ice storage tank, we can further introduce the temperature efficiency weight coefficient γ and the water volume efficiency weight coefficient δ to construct the efficiency function η(t) = γ × T'(t) + δ × V'(t). As an implementation, assume that γ is set to 0.4 and δ is set to 0.6, indicating that the impact of water volume on cooling efficiency is slightly greater than that of temperature. Specifically, at a certain time point t, the water temperature T'(t) is 5°C and the water volume V'(t) is 1000L, then η(t) = 0.4 × 5 + 0.6 × 1000 = 602 units. This efficiency function can initially reflect the cooling capacity of the ice storage tank and provide basic data for subsequent optimization. It should be noted that the values ​​of γ and δ can be determined through experiments or historical data analysis to ensure that they accurately reflect the physical characteristics of the system. Specifically, efficiency weight coefficients are determined through multivariate linear regression of historical cooling data, with the temperature weight γ ranging from 0.32 to 0.45 and the water weight ranging from 0.55 to 0.68. This approach helps improve the scientific nature of efficiency assessment and lays a foundation for optimizing ice-making system operations. Furthermore, the initial cooling efficiency sequence η(t) and historical data are input into a support vector machine model for training, resulting in an efficiency prediction sequence η'(t). For example, assuming the historical data includes minute-by-minute temperature and water volume records for the past week, the support vector machine can learn from these data patterns to predict efficiency trends for the next hour. Specifically, during peak ice-making periods, the model might predict η'(t) to increase from 600 to 650 units, reflecting the efficiency improvement brought about by increased energy input. The advantage of this prediction method is that it can identify efficiency trends in advance, buying time for system adjustments and thus avoiding energy waste or insufficient cooling. The efficiency prediction sequence η'(t) is processed using a sliding window with a window size of 10 minutes and a step size of 1 minute. The mean smoothing method is used to obtain the smoothed efficiency sequence η''(t). As an implementation method, assuming that the data of a certain segment η'(t) is 600, 610, 620, 615, 605, 595, 590, 600, 610, 620, the mean of the first 10 values ​​of the sliding window is (600+610+620+615+605+595+590+600+610+620) / 10=606.5 units. This can effectively reduce the efficiency fluctuations caused by sensor noise and further improve the reliability of the data. Specifically, this smoothing process can help operators more clearly identify the long-term trend of efficiency and avoid misjudging the system status due to short-term fluctuations. The smoothed efficiency sequence η''(t) and the water temperature dynamic sequence T'(t) are linearly regressed using the least squares method to obtain the temperature efficiency correlation coefficient matrix A(t). For example, the analysis found that for every 1°C increase in water temperature, efficiency may increase by 10 units, indicating that temperature has a positive effect on efficiency.As an implementation method, this relationship can be verified through multiple experiments to ensure that the matrix A(t) accurately reflects the contribution of temperature changes. Similarly, a regression analysis is performed on η''(t) and the water volume dynamic sequence V'(t) to obtain the water volume efficiency correlation coefficient matrix B(t). Assuming that every 100L increase in water volume increases efficiency by 15 units, B(t) can quantify this impact. The advantage of this analysis method is that it clearly defines the respective contributions of temperature and water volume to efficiency, providing data support for subsequent optimization. A weighted matrix fusion is performed on the temperature efficiency correlation coefficient matrix A(t) and the water volume efficiency correlation coefficient matrix B(t) to generate the comprehensive efficiency impact matrix C(t). Specifically, the weighting is performed using preset weights γ = 0.4 and δ = 0.6, resulting in C(t) = 0.4 × A(t) + 0.6 × B(t). For example, if the value of an element in A(t) is 10 and the corresponding element in B(t) is 15, then the value of that element in C(t) is 0.4 × 10 + 0.6 × 15 = 13. This fusion method comprehensively considers the synergistic effect of temperature and water volume, making the system status assessment more comprehensive. It should be noted that if the element value in C(t) exceeds the threshold of 5, for example, reaching 13, it indicates that the efficiency is significantly affected and the operation strategy needs to be adjusted. If the element value of C(t) exceeds the threshold of 5, the smoothed efficiency sequence is updated by the differential equation η'''(t)=η''(t-1)+λ×C(t) to obtain the adjusted efficiency sequence η'''(t). As an implementation method, the adjustment coefficient λ is set to 0.1, assuming that η''(t-1) is 600 units and C(t) is 13, then η'''(t)=600+0.1×13=601.3 units. This adjustment mechanism can quickly respond to system changes, such as increasing the efficiency target value when the demand for ice making suddenly increases. Specifically, this method can enhance the adaptability of the system and ensure that the cooling efficiency is always in the best state. Furthermore, continuous monitoring of η'''(t) can provide a basis for equipment maintenance. For example, a long-term decline in efficiency may indicate ice tank aging. For example, in actual operation, during the nighttime off-peak electricity price period, an ice tank was increased in water volume by increasing energy input. This increased T'(t) from 4°C to 6°C, V'(t) from 800L to 1200L, and η(t) from 500 to 730. η'(t) predicted that the next day's peak efficiency would reach 750 units. After smoothing and adjustment, η'''(t) stabilized at 740 units. This multi-step process, from data collection to efficiency optimization, significantly improves energy efficiency, reduces operating costs, and provides reliable support for system prediction and maintenance.

[0029] S105. Obtain building load demand data L(t) and ambient temperature data A(t), use a long short-term memory network to analyze the time series characteristics of L(t), and combine A(t) to calculate the load demand change rate R(t) = |L(t)-L(t-1)| / L(t-1) to determine the dynamic coupling characteristics of the load demand.

[0030] The dynamic sequence of building load demand, Lt, and the dynamic sequence of ambient temperature, At, were obtained. Lt was decomposed into a time series using a long short-term memory network implemented in TensorFlow, yielding the load demand's trend component, T_Lt, and cyclic component, C_Lt. The load demand change rate sequence, Rt, was calculated using the formula Rt=|Lt-Lt-1| / maxLt-1,1, where the denominator is the maximum of Lt-1 and 1 to avoid zero. Combined with the dynamic sequence of ambient temperature, At, the Pearson correlation coefficient from the Scipy library was used to analyze the correlation between Rt and At, yielding the correlation coefficient sequence, Pt. For the correlation coefficient sequence, Pt, a sliding window was applied using the rolling function from the Pandas library, with a window size of 15 minutes and a step size of 2 minutes. Mean smoothing was used to obtain the smoothed correlation coefficient sequence, P't. Based on the smoothed correlation coefficient sequence, a random forest model implemented in Scikit-learn was used to fit the nonlinear relationship between the load demand change rate, Rt, and ambient temperature, At, yielding the fitted relationship function, Ft. If the weights of the variables in the fitted relationship function Ft exceed the preset threshold of 8, the rate of change sequence is updated using the differential equation R't = Rt-1 + μ × Ft, resulting in an adjusted rate of change sequence R't, where μ is the preset adjustment coefficient. The adjusted rate of change sequence R't is predicted using a long short-term memory network implemented in TensorFlow. Combined with the ambient temperature dynamic sequence At, the load demand forecast sequence L't for the next time period is obtained. A weighted fusion method is used based on the load demand forecast sequence L't and the trend component T_Lt, with the weights determined by P't, to obtain the comprehensive load demand dynamic sequence L''t.

[0031] For example, obtaining the dynamic series of building load demand Lt and ambient temperature At can provide key input data for optimizing the operation of ice-making or cooling systems. As an implementation, the dynamic series of building load demand Lt can be understood as the hourly cooling power demand of an office building. For example, it may reach 500kW during the peak period of summer noon and drop to 100kW during the low period of nighttime. The dynamic series of ambient temperature At records the changes in outdoor temperature, for example, from 25°C in the morning to 35°C in the afternoon. Using a long short-term memory network implemented in TensorFlow to perform time series decomposition on Lt, it is possible to extract the trend component T_Lt and the periodic component C_Lt. Specifically, the trend component T_Lt reflects long-term changes in load demand, such as the gradual increase in load during weekdays, while the periodic component C_Lt captures the regular fluctuations during daily peaks in the morning and evening. The benefit of this decomposition is that it breaks down complex load data into interpretable components, laying the foundation for subsequent prediction and optimization. Furthermore, the load demand change rate sequence Rt is calculated using the formula Rt=|Lt-Lt-1| / max(Lt-1,1), which can quantify the dynamic changes in the load. For example, if Lt-1 was 400kW in the previous hour and Lt is currently 450kW, then Rt=|450-400| / 400=0.125, indicating a 12.5% ​​load increase. The denominator takes the maximum value of Lt-1 and 1 to avoid calculation errors when the load is zero. As an implementation method, when the load is extremely low at night, for example, Lt-1 is 0kW and Lt is 50kW, then Rt=50 / 1=50, reflecting the drastic changes caused by the sudden start-up of the equipment. This rate of change sequence can help operators quickly identify the intensity of load fluctuations. Combined with the dynamic ambient temperature sequence At, the correlation between Rt and At is analyzed using the Pearson correlation coefficient in the Scipy library to obtain the correlation coefficient sequence Pt. For example, in a single day's records, when At rises from 30°C to 35°C, Rt increases from 0.1 to 0.15. The calculated Pt might be 0.85, indicating a strong positive correlation between rising temperature and the rate of load change. Specifically, this analysis can reveal how ambient temperature drives electricity demand, providing a basis for energy allocation within the cooling system. It should be noted that the positive or negative value of Pt can also indicate whether temperature changes increase or decrease load fluctuations. For the correlation coefficient sequence Pt, a sliding window is applied using the rolling function in the Pandas library, with a window size of 15 minutes and a step size of 2 minutes. Mean smoothing is then performed to obtain P't. For example, for a period of Pt data with values ​​of 0.8, 0.85, 0.9, 0.87, and 0.82, the first value of P't after smoothing is (0.8 + 0.85 + 0.9 + 0.87 + 0.82) / 5 = 0.848. This smoothing process reduces noise caused by short-term temperature fluctuations and makes the correlation trend clearer. As an implementation method, operators can use this to determine whether the long-term impact of temperature on load is stable.Based on P't, a random forest model implemented in Scikit-learn fits the nonlinear relationship between Rt and At, yielding the fitted relationship function Ft. For example, the model may find that when At exceeds 32°C, the growth rate of Rt accelerates significantly, with the weight potentially reaching 10, exceeding the threshold of 8. Furthermore, if the variable weights in Ft are greater than 8, the rate of change sequence is updated using the difference equation R't = Rt-1 + μ × Ft. Assuming μ is 0.05, Rt-1 is 0.12, and Ft is 10, then R't = 0.12 + 0.05 × 10 = 0.17. This adjustment allows for rapid response to temperature-driven load changes and improves forecast accuracy. Using TensorFlow's long short-term memory network to predict R't and combine it with At, we can derive the load demand forecast sequence L't for the next time period. For example, if At is predicted to be 36°C at noon the next day, L't may increase from the current 480 kW to 520 kW. Specifically, this forecasting method enables pre-planning of ice-making system operating times to avoid cooling shortages. Based on L't and T_Lt, P't is used as the weight for weighted fusion to generate a comprehensive load demand dynamic sequence L''t. For example, if P't is 0.85, T_Lt is 490kW, and L't is 520kW, then L''t=0.85×520+(1-0.85)×490=515.5kW. This fusion method comprehensively considers trends and forecast results, making the load demand assessment more comprehensive. It should be noted that this multi-step analysis, from data collection to forecast optimization, can significantly improve the energy efficiency of the cooling system. Furthermore, continuous monitoring of L''t can also provide clues for equipment maintenance. For example, it is found that a long-term high load forecast may be a signal of reduced ice-making efficiency. The advantage of this method is that it enhances the adaptability and reliability of the system in a data-driven manner.

[0032] S106. Based on the ice melting cooling efficiency η(t) and the load demand change rate R(t), a dynamic coupling relationship model C(t)=ω×η(t)+ψ×R(t) is constructed, where ω and ψ are preset coupling weights. If C(t) exceeds the preset threshold, the time lag effect characteristic τ(t)=argmax(C(t)-C(t-τ)) is calculated using the sliding window method to determine the energy transfer delay between ice making and cooling.

[0033] By collecting time series data on the ice-melting cooling efficiency η(t) and the load demand change rate R(t), the dynamic coupling relationship sequence is calculated using the formula C(t) = ω × η(t) + ψ × R(t), where ω and ψ are preset coupling weights. This yields the coupling model sequence C(t). If C(t) exceeds a preset threshold, the C(t) sequence is processed using a sliding window method with a 20-minute window size and a 5-minute step size. The time lag effect characteristic τ(t) = argmax_{τ∈[5,30]}(C(t) - C(t - τ)) is calculated to determine the time delay τ, resulting in the time lag characteristic sequence τ(t), where τ is expressed in minutes. The support vector machine algorithm in Scikit-learn is used, with the radial basis function as the kernel function. C(t) and τ(t) are used as inputs for nonlinear mapping, resulting in the mapping relationship function M(t). A long short-term memory network implemented in TensorFlow takes the mapping function M(t) and the ice-melting cooling efficiency η(t) as inputs, performs joint time series analysis, and predicts the coupling model sequence C'(t) for the next time period. Based on the predicted C'(t) and the time-delay characteristic sequence τ(t), a weighted fusion method is used to calculate the adjusted coupling sequence C'(t). The weight is determined by the product of M(t) and C'(t), resulting in the adjusted coupling sequence C'(t). Based on the adjusted coupling sequence C'(t), the load demand change rate is updated using the differential equation R'(t)=R(t-1)+λ×C'(t), where λ is the preset adjustment coefficient. This results in the adjusted change rate sequence R'(t). A random forest model implemented in Scikit-learn takes the adjusted change rate sequence R'(t) and the time delay τ as inputs, performs regression analysis, and obtains the fitted relationship function F(t) between the load demand change rate and the energy transfer delay, thereby determining the final dynamic coupling characteristics.

[0034] For example, by collecting time series data on the ice-melting cooling efficiency η(t) and the load demand change rate R(t), a foundation for dynamic optimization of the cooling system can be provided. For example, in an office building's ice-making cooling system, η(t) represents the energy conversion efficiency of the ice-melting process. It may reach 0.9 during ice-making at night, when the load is low, and drop to 0.75 during peak daytime hours due to increased equipment load. R(t) reflects the rate of change in load demand, such as a surge from 0.1 to 0.15 at noon due to air conditioning usage. The dynamic coupling relationship sequence is calculated using the formula C(t) = ω × η(t) + ψ × R(t). Assuming ω is 0.6 and ψ is 0.4, then at a certain moment when η(t) is 0.8 and R(t) is 0.12, C(t) = 0.6 × 0.8 + 0.4 × 0.12 = 0.528. As an implementation, C(t) quantifies the combined impact of efficiency and load changes. If it exceeds a preset threshold of 0.5, it indicates that the system requires dynamic adjustment. Specifically, when C(t) exceeds the threshold, the C(t) sequence is processed using a sliding window method with a window size of 20 minutes and a step size of 5 minutes to calculate the time lag effect feature τ(t). For example, on a certain day, C(t) slowly rises from 0.52 to 0.58. By comparing the data before and after, it is found that the maximum difference occurs 15 minutes ago, so τ(t) is 15 minutes. This reflects that there is a delay in the impact of load changes on cooling efficiency. Furthermore, using the support vector machine algorithm, with C(t) and τ(t) as inputs and selecting the radial basis function kernel, it can capture nonlinear relationships. For example, when τ(t) is 10 minutes and C(t) is 0.55, the mapping function M(t) may output 0.62, indicating the potential correlation strength between the two. The benefit of this method is that it reveals how time delay affects system response. As an implementation method, M(t) and η(t) are analyzed through a long short-term memory network to predict C'(t) for the next time period. For example, if M(t) is currently 0.6 and η(t) is 0.78, the network might predict C'(t) to be 0.59. This allows for early assessment of the system coupling state, helping operators adjust ice-making strategies. Combining C'(t) and τ(t), a weighted fusion method is used to calculate C'(t), with the weight determined by the product of M(t) and C'(t). For example, if M(t) is 0.6 and C'(t) is 0.59, the weight is 0.354, and the fused C'(t) might be 0.57. This adjustment makes the prediction more accurate to the actual operating state, improving reliability. Furthermore, the rate of change is updated using the difference equation R'(t) = R(t-1) + λ × C''(t). Assuming λ is 0.05, R(t-1) is 0.13, and C''(t) is 0.57, then R'(t) = 0.13 + 0.05 × 0.57 = 0.1585. This reflects the dynamic impact of the coupling effect on load changes. It should be noted that this update can quickly respond to system changes and avoid cooling shortages.For example, at noon one day, the temperature suddenly rose, and R'(t) increased from 0.12 to 0.16, indicating that the amount of ice melting needed to be increased. Specifically, the random forest model was used to analyze R'(t) and τ(t) to obtain the fitting relationship function F(t). For example, when τ(t) is 12 minutes and R'(t) is 0.15, F(t) may output 0.18, indicating that the delay is positively correlated with the rate of change. This can help operators understand the impact of energy transfer delay on the load. For example, during the peak summer season, when τ(t) is long, F(t) prompts that the cooling equipment needs to be started in advance. As an implementation method, if F(t) shows that τ(t) exceeds 15 minutes and R'(t) increases rapidly, the amount of ice production can be increased 20 minutes in advance to ensure stable cooling. This multi-step analysis progresses from efficiency, rate of change to time lag characteristics to ensure accurate energy allocation.

[0035] S107. Extract the full-cycle features from the dynamic coupling relationship C(t) and the time lag effect τ(t), and use the random forest algorithm to perform multivariate interactive regression analysis on T'(t), V'(t), η(t), and R(t) to obtain the energy efficiency prediction value Q(t) of the system at different time scales.

[0036] By dynamically coupling full-cycle data acquisition C(t), time lag effect τ(t), ice melting efficiency η(t), and load variation R(t), a preprocessing method is used to extract the periodic features T'(t) and velocity features V'(t) to obtain an initial feature sequence. Based on this initial feature sequence, a random forest algorithm is used to perform multivariate regression on the periodic features T'(t), velocity features V'(t), ice melting efficiency η(t), and load variation R(t), resulting in a preliminary energy efficiency prediction sequence Q1(t). For this preliminary energy efficiency prediction sequence Q1(t), a sliding window method is used to calculate the fluctuation feature sequence F(t) of Q1(t) within the time scale, with a window size of 30 minutes and a step size of 10 minutes. This results in the fluctuation feature sequence F(t). If the fluctuation feature sequence F(t) exceeds a preset threshold, a support vector machine algorithm is used to perform nonlinear mapping between F(t) and the time lag effect τ(t), using the radial basis function as the kernel function, to obtain the mapped feature sequence M(t). A long short-term memory network (LSTM) is used to perform a joint time series analysis of the mapping feature sequence M(t), dynamic coupling C(t), and ice melting efficiency η(t) to predict the energy efficiency prediction sequence Q2(t) for the next time period. Based on the predicted energy efficiency prediction sequence Q2(t) and the fluctuation feature sequence F(t), a weighted fusion method is used to calculate the adjusted energy efficiency sequence Q3(t). The weight is determined by the product of M(t) and Q2(t), resulting in the adjusted energy efficiency sequence Q3(t). The load change sequence R'(t) = R(t-1) + λ × Q3(t) is updated using a differential equation, where λ is a preset adjustment coefficient. A random forest algorithm is used to perform regression analysis on R'(t), the time lag effect τ(t), and the time scale to obtain the final energy efficiency prediction sequence Q(t).

[0037] For example, dynamic coupling of full-cycle data collection, C(t), time lag effect τ(t), ice melting efficiency η(t), and load variation R(t), can provide a data foundation for dynamic optimization of building energy systems. As an implementation, a preprocessing method extracts periodic features T'(t) and velocity features V'(t) to capture the cyclical patterns and rate of change of system operation. For example, in an office building's ice-making and cooling system, 24-hour data collection revealed that η(t) was higher at night, approximately 0.9, due to peak ice-making efficiency during low load periods. During the day, η(t) dropped to 0.7 due to increased air conditioning load. T'(t) may indicate a significant 24-hour cycle, while V'(t) reflects the rate of decrease in η(t) from night to day, such as a decrease of 0.05 per hour. This feature extraction lays the foundation for subsequent analysis and helps reveal the dynamic relationship between efficiency and load. Specifically, based on the initial feature sequence, the random forest algorithm performs multivariate regression on T'(t), V'(t), η(t), and R(t) to generate a preliminary energy efficiency prediction sequence Q1(t). The advantage of random forests lies in their ability to handle complex relationships between multiple variables. For example, during the summer peak season, T'(t) indicates that load demand fluctuates every 12 hours, V'(t) indicates an accelerated rate of change, R(t) increases from 0.1 to 0.15, and η(t) is 0.75. Taking these factors into account, the random forest algorithm may predict Q1(t) to be 0.65, representing the current energy efficiency status. This step provides a preliminary prediction, laying the foundation for subsequent refined analysis. Furthermore, a sliding window method is used to calculate the fluctuation feature sequence F(t) for Q1(t), with a window size of 30 minutes and a step size of 10 minutes. This is used to quantify short-term fluctuations in energy efficiency. For example, at noon one day, Q1(t) increased from 0.62 to 0.68. The calculated F(t) indicates a fluctuation amplitude of 0.06. If F(t) exceeds the preset threshold of 0.05, it indicates significant fluctuations in system energy efficiency and requires further action. This reflects the impact of sudden load changes on energy efficiency and helps operators detect anomalies promptly. As an implementation, if F(t) exceeds the threshold, a support vector machine (SVM) uses a radial basis function kernel to perform nonlinear mapping between F(t) and τ(t) to generate a mapping feature sequence M(t). For example, if τ(t) is 15 minutes and F(t) is 0.06, M(t) might output 0.7, revealing a potential correlation between fluctuations and time lags. This method captures nonlinear characteristics, improves prediction accuracy, and supports subsequent time series analysis. A long short-term memory (LSTM) network jointly analyzes M(t), C(t), and η(t) to predict the energy efficiency sequence Q2(t) for the next time period. Specifically, assuming the current M(t) is 0.7, C(t) is 0.58, and η(t) is 0.76, the network might predict Q2(t) to be 0.61. This step utilizes the time dependency of historical data to predict future energy efficiency trends, which helps to adjust ice-making strategies in advance and avoid insufficient cooling.Based on Q2(t) and F(t), a weighted fusion is performed to calculate the adjusted energy efficiency sequence Q3(t), with the weight determined by the product of M(t) and Q2(t). For example, if M(t) is 0.7 and Q2(t) is 0.61, the weight is 0.427, and Q3(t) may be 0.6 after fusion. This adjustment makes the prediction closer to actual operating conditions and improves reliability, especially in scenarios with large load fluctuations. Furthermore, the load change sequence is updated using the difference equation R'(t) = R(t-1) + λ × Q3(t), with λ set to 0.05. For example, if R(t-1) is 0.13 and Q3(t) is 0.6, R'(t) is updated to 0.16. This reflects the dynamic impact of energy efficiency on load changes, ensuring that the system responds quickly to changes in demand. It should be noted that random forest regression analysis is performed on R'(t), τ(t), and time scale to generate the final energy efficiency prediction sequence Q(t). For example, if τ(t) is 12 minutes and R'(t) is 0.15, Q(t) might be 0.67. This multivariate regression approach comprehensively accounts for the impact of time lag and load variations. For example, on hot summer days, τ(t) is long and R'(t) increases rapidly, while Q(t) suggests the need to start equipment early to ensure stable cooling. This progressive analysis, from feature extraction to predictive optimization, not only reduces energy consumption but also improves system adaptability, providing a scientific basis for energy management.

[0038] S108. Obtain the prediction result of Q(t). For historical data under extreme weather scenarios, if the deviation between Q(t) and the actual energy efficiency Q'(t) |D(t)|=|Q(t)-Q'(t)| exceeds the preset threshold, adjust the model parameters through adaptive Kalman filtering and update the full-cycle energy efficiency evaluation framework.

[0039] The predicted energy efficiency sequence Q(t) and the actual energy efficiency sequence Q'(t) are obtained from historical data under extreme weather scenarios. The deviation sequence |D(t)| = |Q(t) - Q'(t)| is calculated, and the distribution characteristics of the deviation sequence |D(t)| are obtained. If the deviation sequence |D(t)| exceeds a preset threshold, an adaptive Kalman filter is used to adjust the model parameters. The full-cycle feature sequence T'(t) from the historical data is combined with the dynamic coupling relationship C(t) to obtain the updated model parameter sequence P(t). Based on the updated model parameter sequence P(t), the nonlinear mapping characteristics between the time lag effect sequence τ(t) and the load change sequence R(t) are recalculated. The full-cycle feature sequence T'(t) is processed using the support vector machine algorithm, and the radial basis function is selected as the kernel function to obtain the mapping feature sequence M'(t). Based on the mapping feature sequence M'(t) and the dynamic coupling relationship C(t), a long short-term memory network is used to perform time series prediction on the time lag effect sequence τ(t), obtaining the time lag effect prediction sequence τ'(t) for the next time period. The time-lag effect prediction sequence τ'(t) and the load change sequence R(t) are obtained. A multivariate regression analysis is performed on the full-cycle feature sequence T'(t), the mapping feature sequence M'(t), and the time-lag effect prediction sequence τ'(t) using the random forest algorithm. This yields the adjusted load change prediction sequence R'(t). The time-scale fluctuation feature sequence F'(t) is calculated using a sliding window method with a 30-minute window size and a 10-minute step size for the adjusted load change prediction sequence R'(t). The final energy efficiency prediction sequence Q'(t) is calculated using a weighted fusion method based on the fluctuation feature sequence F'(t) and the adjusted load change prediction sequence R'(t). The weights are determined by multiplying the mapping feature sequence M'(t) and the time-lag effect prediction sequence τ'(t). This yields the final energy efficiency prediction sequence Q'(t).

[0040] For example, the predicted energy efficiency sequence Q(t) and the actual energy efficiency sequence Q'(t) are obtained through historical data under extreme weather scenarios, and the deviation sequence |D(t)|=|Q(t)-Q'(t)| is calculated to evaluate the prediction accuracy of the model under abnormal conditions. For example, in a city’s summer typhoon weather, historical data records the operating status of the building’s cooling system. The predicted energy efficiency Q(t) is 0.68, while the actual energy efficiency Q'(t) is 0.62, and the deviation |D(t)| is 0.06. Furthermore, by statistically analyzing data from multiple days, the deviation sequence |D(t)| shows a normal distribution characteristic, with a mean of approximately 0.05 and a standard deviation of 0.02. If the preset threshold is 0.04, the deviation in some time periods exceeds the threshold, indicating that the model prediction is deviated under extreme weather conditions and needs to be optimized. As an implementation method, when the deviation sequence |D(t)| exceeds a threshold, the adaptive Kalman filter dynamically adjusts model parameters, combining the full-cycle characteristic sequence T'(t) and the dynamic coupling relationship C(t) to generate an updated parameter sequence P(t). Specifically, in the aforementioned typhoon scenario, T'(t) indicates increased load fluctuations within a 24-hour period, while C(t) reflects a decrease in the coupling strength between devices from 0.55 to 0.45. Based on these characteristics, the filtering algorithm adjusts parameters to make P(t) more adaptable to dynamic changes under extreme conditions. This approach improves the model's robustness to sudden changes in the environment, for example, by preventing inaccurate predictions caused by sudden load increases. It should be noted that the updated parameter sequence P(t) is used to recalculate the nonlinear mapping features of the time-lag effect sequence τ(t) and the load variation sequence R(t). A support vector machine processes T'(t) with a radial basis function kernel to generate the mapping feature sequence M'(t). For example, during a heavy rainstorm at a hospital, T'(t) exhibited a 12-hour period, while τ(t) extended from 10 minutes to 15 minutes. The support vector machine output, M'(t), was 0.72, revealing a nonlinear relationship between time lag and period. This mapping can capture complex relationships and improve the accuracy of subsequent forecasts. Furthermore, using M'(t) and C(t), a long short-term memory network (LSTM) performs time series forecasts on τ(t), generating a time lag effect forecast sequence, τ'(t), for the next time period. For example, during a heavy snowstorm, M'(t) was 0.68 and C(t) was 0.50, and the network predicted that τ'(t) would increase from 12 minutes to 18 minutes. This forecast leverages historical time dependence and helps to proactively adjust cooling strategies to avoid energy efficiency degradation due to extended time lags. Specifically, after obtaining τ'(t) and R(t), a random forest algorithm performs multivariate regression on T'(t), M'(t), and τ'(t) to generate an adjusted load change forecast sequence, R'(t). For example, during hot and humid weather, T'(t) exhibits a six-hour fluctuation, M'(t) is 0.70, τ'(t) is 14 minutes, and R(t) increases from 0.12 to 0.16. The random forest prediction R'(t) is 0.17. This approach integrates the influence of multiple variables to ensure that load forecasts closely match actual operating conditions.As an implementation, a sliding window method is used to calculate the fluctuation characteristic sequence F'(t) for R'(t) using a 30-minute window and a 10-minute step size. For example, during a thunderstorm, R'(t) for a commercial building rises from 0.15 to 0.18, and the calculated fluctuation amplitude of F'(t) is 0.03. If the threshold is 0.05, the fluctuation is controllable; if it exceeds the threshold, the operator is prompted to pay attention to the sudden load change. This method of quantifying short-term fluctuations helps to promptly detect anomalies and optimize system response. Furthermore, based on F'(t) and R'(t), a weighted fusion method is used to calculate the final energy efficiency prediction sequence Q'(t), with the weight determined by the product of M'(t) and τ'(t). For example, if M'(t) is 0.72 and τ'(t) is 15 minutes, the weight is 0.72 × 15 / 60 = 0.18, and the fused Q'(t) is 0.65. This adjustment makes the prediction more closely aligned with actual operating conditions, especially in extreme weather conditions. For example, during a snowstorm, Q'(t) indicated that energy efficiency in an office building had dropped to 0.60. Operators proactively increased ice production to ensure stable cooling. This progressive analysis, from deviation assessment to parameter optimization and then to multivariate prediction, not only improves prediction accuracy but also enhances the system's adaptability to extreme conditions, providing reliable support for energy management.

[0041] Also included is an energy efficiency evaluation and optimization system for a photovoltaic power generation and cold storage system for implementing any one of the technical solutions described in steps S101-S108, characterized by comprising:

[0042] A data acquisition module is used to obtain real-time power data of the photovoltaic power station, weather conditions, ice storage tank temperature, water volume, building load demand and ambient temperature data through a sensor network;

[0043] The power fluctuation analysis module is used to calculate the power fluctuation sequence based on the sliding window algorithm and the Pearson correlation coefficient to generate the power fluctuation quantitative results;

[0044] Dynamic modeling module, used to process smooth fluctuation feature sequences through convolutional neural networks and random forest algorithms, generate real-time energy input models, and output dynamic change parameters of temperature and water series;

[0045] The coupling relationship construction module is used to integrate the support vector machine model and the long short-term memory network to process the time lag effect of the cooling efficiency sequence and the load demand change rate sequence, and generate a dynamic coupling relationship sequence and an energy transfer delay sequence;

[0046] Energy efficiency prediction engine module, which is used to perform multivariate regression analysis and fluctuation feature mapping processing, and generate a full-cycle energy efficiency prediction sequence based on the comprehensive dynamic impact matrix and period feature extraction;

[0047] The feedback optimization module is used to adaptively calibrate the influence coefficient sequence and dynamic coupling relationship sequence according to the preset threshold, and update the system evaluation parameters through the matrix weighted fusion algorithm.

[0048] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for energy efficiency evaluation and optimization of a photovoltaic power generation and cold storage system, characterized in that: include: Obtain real-time power data and weather condition data from photovoltaic power stations and calculate power fluctuation rate series; Extracting characteristic values ​​from the power fluctuation rate sequence to determine a sequence of influence coefficients of the fluctuation on ice-making energy input; Acquire temperature data and water volume data in the ice storage tank, update the real-time energy input model through the influence coefficient sequence, and calculate the dynamic changes of the temperature sequence and the water volume sequence; Constructing a cooling efficiency sequence according to the dynamic changes of the temperature sequence and the water quantity sequence; Obtain building load demand data and ambient temperature data, and calculate the load demand change rate sequence; By using the cooling efficiency sequence and the load demand change rate sequence, a dynamic coupling relationship sequence is constructed to determine an energy transfer delay sequence; Extracting full-cycle features from the dynamic coupling relationship sequence and the energy transfer delay sequence to generate an energy efficiency prediction sequence; The step of obtaining real-time power data and weather condition data of a photovoltaic power station and calculating a power fluctuation rate sequence includes: The photovoltaic output power and sunshine intensity are collected by sensors, and the power fluctuation rate is calculated using the time series analysis method. If the power at the previous time point is zero and the current power is non-zero, the power fluctuation rate is set to the preset value; The power fluctuation rate and sunshine intensity are processed by the sliding window method, and the average power fluctuation rate and average sunshine intensity in continuous time periods are calculated. The Pearson correlation coefficient is used to analyze the correlation between the average power fluctuation rate and the average sunshine intensity, and the quantitative results of power fluctuation are obtained.

2. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The step of extracting characteristic values ​​from the power fluctuation rate sequence and determining a sequence of influence coefficients of the fluctuation on ice-making energy input includes: The power fluctuation sequence is processed by sliding window method to extract the smooth fluctuation feature sequence; The energy input per unit time is calculated by integration, and the instantaneous mapping relationship between the smooth fluctuation feature sequence and the energy input is analyzed using a convolutional neural network to obtain a mapping feature matrix. The influence coefficient sequence is calculated based on the mapping feature matrix. If the influence coefficient exceeds a preset threshold, the relationship between the smooth fluctuation feature sequence and the energy input is processed using a random forest algorithm, and the influence coefficient sequence is adjusted to obtain a dynamic influence sequence of the fluctuation on the energy input.

3. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The updating of the real-time energy input model by using the influence coefficient sequence and calculating the dynamic changes of the temperature sequence and the water volume sequence include: The energy sequence is updated through the influence coefficient sequence and energy input calculation, and the temperature sequence and water volume sequence are generated using the difference equation; The temperature series and water volume series are processed by sliding window method to obtain smooth temperature series and smooth water volume series. The mapping relationship between the smoothed temperature series and the updated energy series is analyzed by linear regression to obtain the temperature-energy correlation coefficient matrix. The temperature energy correlation coefficient matrix and the water volume energy correlation coefficient matrix are combined through a matrix fusion method to generate a comprehensive dynamic influence matrix, and the temperature sequence and the water volume sequence are adjusted according to the comprehensive dynamic influence matrix.

4. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The step of constructing a cooling efficiency sequence according to the dynamic changes of the temperature sequence and the water quantity sequence includes: An initial cooling efficiency sequence is calculated using a preset efficiency weight coefficient. The initial cooling efficiency sequence and historical data are input into a support vector machine model to generate an efficiency prediction sequence. The efficiency weight coefficient is determined by multivariate linear regression of historical cooling data, where the temperature weight γ ranges from 0.32 to 0.45 and the water weight ranges from 0.55 to 0.

68. The efficiency prediction sequence is processed by the sliding window method to obtain a smooth efficiency sequence; The least square method is used to analyze the linear relationship between the smoothed efficiency series and the temperature series, and the temperature efficiency correlation coefficient matrix is ​​obtained; The temperature efficiency correlation coefficient matrix and the water efficiency correlation coefficient matrix are fused by matrix weighting to generate a comprehensive efficiency impact matrix, and the smoothed efficiency sequence is adjusted according to the comprehensive efficiency impact matrix.

5. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The step of obtaining building load demand data and ambient temperature data and calculating a load demand change rate sequence includes: Decomposing load demand data through long short-term memory network, obtaining trend component and period component, and calculating load demand change rate sequence; The Pearson correlation coefficient is used to analyze the correlation between the load demand change rate series and the ambient temperature data, and the correlation coefficient series is obtained; The correlation coefficient sequence is processed by the sliding window method to obtain a smoothed correlation coefficient sequence; A random forest model is used to fit the nonlinear relationship between the smoothed correlation coefficient sequence and the ambient temperature data to generate a fitting relationship function, and the load demand change rate sequence is adjusted according to the fitting relationship function.

6. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The step of constructing a dynamic coupling relationship sequence and determining an energy transfer delay sequence by using the cooling efficiency sequence and the load demand change rate sequence includes: The dynamic coupling relationship sequence is calculated using the preset coupling weight. If the dynamic coupling relationship sequence exceeds the preset threshold, the time lag effect characteristics are calculated using the sliding window method to generate an energy transfer delay sequence. The support vector machine algorithm is used to process the dynamic coupling relationship sequence and the energy transfer delay sequence to obtain the mapping relationship function; The mapping relationship function and the cooling efficiency sequence are analyzed through the long short-term memory network, the dynamic coupling relationship sequence of the next time period is predicted, and the dynamic coupling relationship sequence is adjusted according to the prediction results.

7. The energy efficiency evaluation and optimization method of the photovoltaic power generation and cold storage system according to claim 1, characterized in that: The extracting full-cycle features from the dynamic coupling relationship sequence and the energy transfer delay sequence to generate an energy efficiency prediction sequence includes: The periodic characteristics of the dynamic coupling relationship sequence and the energy transfer delay sequence were extracted through preprocessing methods. The random forest algorithm was used to perform multivariate regression on the temperature sequence, water volume sequence, cooling efficiency sequence, and load demand change rate sequence to generate a preliminary energy efficiency prediction sequence. The fluctuation characteristic sequence of the preliminary energy efficiency prediction sequence is calculated by the sliding window method. If the fluctuation characteristic sequence exceeds a preset threshold, the support vector machine algorithm is used to process the fluctuation characteristic sequence and the energy transfer delay sequence to obtain a mapping characteristic sequence, and the energy efficiency prediction sequence is adjusted according to the mapping characteristic sequence.

8. An energy efficiency evaluation and optimization system for a photovoltaic power generation and cold storage system for implementing the method according to any one of claims 1 to 7, characterized in that: Includes: A data acquisition module is used to obtain real-time power data of the photovoltaic power station, weather conditions, ice storage tank temperature, water volume, building load demand and ambient temperature data through a sensor network; The power fluctuation analysis module is used to calculate the power fluctuation sequence based on the sliding window algorithm and the Pearson correlation coefficient to generate the power fluctuation quantitative results; Dynamic modeling module, used to process smooth fluctuation feature sequences through convolutional neural networks and random forest algorithms, generate real-time energy input models, and output dynamic change parameters of temperature and water series; The coupling relationship construction module is used to integrate the support vector machine model and the long short-term memory network to process the time lag effect of the cooling efficiency sequence and the load demand change rate sequence, and generate a dynamic coupling relationship sequence and an energy transfer delay sequence; Energy efficiency prediction engine module, which is used to perform multivariate regression analysis and fluctuation feature mapping processing, and generate a full-cycle energy efficiency prediction sequence based on the comprehensive dynamic impact matrix and period feature extraction; The feedback optimization module is used to adaptively calibrate the influence coefficient sequence and dynamic coupling relationship sequence according to the preset threshold, and update the system evaluation parameters through the matrix weighted fusion algorithm.

Images