Integrated Intelligent Energy

Multiscale convolution-residual network for short-term wind power forecasting

YIN Linfei, TONG Bowen, LI Wenji

2025, 47(6): 1-10. doi:10.3969/j.issn.2097-0706.2025.06.001

Asbtract ( 95 )

HTML ( 2)

PDF (2304KB) ( 39 )

Figures and Tables | References | Related Articles | Metrics

Short-term wind power forecast can provide a basis for power system scheduling and prevent power systems from severe impacts of wind power fluctuations. To improve the low prediction accuracy of existing deep learning models applied for short-term wind power prediction，a short-term wind power prediction method based on multiscale convolution-residual network is proposed to optimize the models. The proposed multiscale convolution-residual network is characterized by full range of feature extraction scales and strong stability，and the multiscale convolution part taking layers with convolutional kernel sizes of 3×3，5×5，7×7 and 9×9 is used to extracted detailed information and global information from the input data. By introducing hopping connections to the residual block，the vanishing gradient problem in the convolutional neural network is solved. The results of the simulation applying on the Natal 378-day dataset show that，the multiscale convolution-residual network can make an accurate prediction on wind power for the next 24 h，and the mean square error of the proposed network is more than 43.55% smaller than that of DarkNet19，InceptionResNetV2，InceptionV3，ResNet18，ResNet50，ShuffleNet and Xception.

Photovoltaic power output prediction based on variational mode decomposition and triple convolutional neural networks

YIN Linfei, ZHANG Yiling

2025, 47(6): 11-19. doi:10.3969/j.issn.2097-0706.2025.06.002

Asbtract ( 61 )

HTML ( 2)

PDF (2236KB) ( 17 )

Figures and Tables | References | Related Articles | Metrics

To address the issue of low accuracy in photovoltaic power output prediction，this paper proposes a prediction model combining variational mode decomposition with triple convolutional neural networks（VMD-TCNNs）. The variational mode decomposition（VMD） was adopted to decompose daily meteorological data，effectively separating intrinsic mode functions. The functions were then stitched，reconstructed and compressed into four-dimensional images，which were input into triple convolutional neural networks（TCNNs） for training and prediction. The initial prediction results from the TCNNs were further processed through fully connected layer，the dropout layer，and the regression output layer to obtain the final results. The VMD-TCNNs model was applied to predict the hourly photovoltaic power output one day in advance in Natal，Brazil. Simulations using actual photovoltaic output and meteorological data were conducted，and the results were compared with 26 other algorithms. The experimental results showed that the mean absolute error （MAE），mean square error and root mean square error of the VMD-TCNNs model were 49.05 W，7403.94 W² and 86.05 W，respectively. Compared with other models，the evaluation indexes of the VMD-TCNNs model were lower，and its MAE value was at least 33.074% smaller than that of the other 26 models，confirming the validity of the proposed model.

Short-term wind power prediction based on VMD-BP-BiLSTM

CHENG Xianlong, ZHANG Jie, LI Siying, YANG Yixia, YANG Cuifei

2025, 47(6): 20-29. doi:10.3969/j.issn.2097-0706.2025.06.003

Asbtract ( 66 )

HTML ( 5)

PDF (2082KB) ( 30 )

Figures and Tables | References | Related Articles | Metrics

With the continuous development of the green energy concept，wind power generation has become a research focus due to its renewable and non-polluting characteristics. However，the output of wind turbines exhibits significant volatility and randomness，posing challenges for power dispatch in the grid. To accurately predict wind power and achieve supply-demand balance and stable operation of the power grid，an innovative variational mode decomposition-back-propagation-bidirectional long short-term memory（VMD-BP-BiLSTM） combined model was proposed as a prediction tool. This model used the average values of adjacent data to detect and replaced outliers in the raw data，followed by data normalization to reduce differences and interference between different data sets. After data preprocessing，VMD was applied to decompose historical wind power generation data into multiple modal components with different characteristics. These modal components，along with corresponding meteorological data，were then input into a combined model of BP neural network and BiLSTM model to independently predict each component. Simulation tests of wind power prediction for wind farms in the northwest region showed that，compared to traditional models such as BP neural networks，BiLSTM，extreme learning machine （ELM），and convolutional neural network-long short-term memory（CNN-LSTM） models，the VMD-BP-BiLSTM model demonstrated more accurate prediction ability. The VMD-BP-BiLSTM combined model provides a new approach for wind power prediction.

A prediction method for power grid carbon emission factor based on T-Graphormer

ZHAN Guohua, ZHANG Xianyong, WEI Shengying, ZHANG Xiaoshun, LI Li

2025, 47(6): 30-36. doi:10.3969/j.issn.2097-0706.2025.06.004

Asbtract ( 57 )

HTML ( 9)

PDF (1871KB) ( 18 )

Figures and Tables | References | Related Articles | Metrics

The carbon emission factor of the power grid is an important indicator for assessing the environmental impact of electricity consumption. Accurate prediction of the power grid carbon emission factor for future time periods is crucial for guiding users to actively participate in demand-side response and achieving clean and low-carbon electricity utilization. Based on the typical spatio-temporal fusion characteristics of the power grid's energy flow，a prediction model for hourly power grid carbon emission factor is proposed，utilizing the T-Graphormer graph neural network. The model incorporates topological information from power grid nodes and historical carbon emission factor data. Through a gated temporal convolution block，the carbon emission factor is mapped into a high-dimensional space，with central and positional encodings embedded into node features. An encoder-decoder structure is then employed for spatio-temporal data mining，and the predicted power grid carbon emission factor is obtained through a multi-layer perceptron. The performance of the proposed model is validated using carbon emission factor data from regions of the UK national grid. The results demonstrates that the prediction model outperforms traditional graph neural network prediction models.

Research on economic scheduling of power systems with wind farms based on improved African vulture optimization algorithm

CHENG Xianlong, MA Yun, HAN Junfeng, MO Ying, GAO Yan

2025, 47(6): 37-46. doi:10.3969/j.issn.2097-0706.2025.06.005

Asbtract ( 46 )

HTML ( 2)

PDF (1933KB) ( 25 )

Figures and Tables | References | Related Articles | Metrics

Due to the increasingly prominent non-renewability and polluting nature of traditional fossil fuels，the research and application of new types of clean energy have become more extensive and in-depth，with wind power generation accounting for a growing share in power systems. However，the intermittency and randomness of wind energy increase the difficulty of solving the economic scheduling problem in power systems with wind farms. To address this complex issue，an economic scheduling model for power systems with wind farms was established，considering both economic and environmental costs. The model aimed to improve both the economic efficiency and environmental sustainability of power grid scheduling，incorporating system load-power balance and unit output constraints as key constraints. Moreover，a multi-objective improved African vulture optimization algorithm was proposed，which integrated Tent chaos mapping and adaptive weight strategies to effectively tackle complex scheduling problems. Simulation experiments were conducted on the modified IEEE 30 system under different objective functions and operation statuses. Taking the low wind power penetration and low load scenario as an example，the compromise solution scores using the multi-objective improved African vulture optimization algorithm improved by 59.105 6%，88.451 8%，37.349 2%，10.147 7%，and 12.700 3% compared to the benchmark algorithms multi-objective particle swarm optimization，non-dominated sorting genetic algorithm，multi-objective grey wolf optimization，multi-objective atomic orbital search，and multi-objective African vulture optimization algorithm，respectively. the simulation results confirmed the feasibility and applicability of the proposed economic scheduling model and multi-objective improved Arican vulture optimization algorithm in real-world power systems.

Remaining useful life prediction of proton exchange membrane fuel cells based on improved HHO-LSTM-Self-Attention

JIANG Jian, DU Dongsheng, SU Lin

2025, 47(6): 47-56. doi:10.3969/j.issn.2097-0706.2025.06.006

Asbtract ( 54 )

HTML ( 0)

PDF (1905KB) ( 29 )

Figures and Tables | References | Related Articles | Metrics

Proton exchange membrane fuel cells（PEMFCs） are widely used in various fields. However，their performance degradation can reduce power output and energy conversion efficiency，and shorten service life. Accurate remaining useful life（RUL） prediction of PEMFCs is crucial for system maintenance，cost reduction，and stable power supply. Based on the temporal variation trend of PEMFC power output，a RUL prediction model that integrated improved Harris Hawks Optimization（HHO） algorithm，long short-term memory（LSTM） network，and self-attention mechanism was proposed. The time-power variation curve was derived from the relationship between current and voltage data. A combination of wavelet adaptive denoising and exponential smoothing was used for decomposition，denoising，and reconstruction of time-power data. To address issues such as excessive training parameters and high computational cost of LSTM，a method combining logistic chaotic mapping with the HHO algorithm was proposed to optimize LSTM，improving training speed and prediction accuracy. Leveraging the self-attention mechanism's advantages in focusing on key information and enhancing training accuracy，the HHO-LSTM-Self-Attention prediction model was established. Experimental results showed that compared with other prediction models such as HHO-LSTM，LSTM，Sparrow Search Algorithm（SSA）-LSTM，and Particle Swarm Optimization （PSO）-LSTM，the proposed model achieved higher prediction accuracy.

Fuzzy active disturbance rejection control of PEMFC air intake unit based on snake optimization algorithm

LI Xiaoning, SUN Na, HUANG Amin, DONG Haiying

2025, 47(6): 57-73. doi:10.3969/j.issn.2097-0706.2025.06.007

Asbtract ( 51 )

HTML ( 1)

PDF (4309KB) ( 19 )

Figures and Tables | References | Related Articles | Metrics

The strong coupling between air flow and pressure in proton exchange membrane fuel cells （PEMFC） results in poor output stability and slow response speed of cell systems. To address these issues，an improved active disturbance rejection control （ADRC） strategy incorporating snake optimization algorithm and fuzzy control was proposed based on the feedforward compensation decoupling structure of the air supply system. The state-space equations of the air supply system were established based on the feedforward compensation decoupling structure. A coupling matrix between flow and pressure was derived which was designed to eliminate the coupling between air flow and pressure. For the improved ADRC design，the feedback control law gain was divided into two parts： the first part was coarsely adjusted using fuzzy control，and the parameters for the second part of state observer gain and feedback control law gain were precisely tuned using the improved snake optimization algorithm. The improved snake optimization algorithm incorporated the chaotic mapping strategy using sequence-based particle swarm optimization method （SPM） for population initialization，enhancing population diversity. Dynamic inertia weights and triangular walk strategy addressed limitations such as slow optimization speed during the early exploration phase. Lens imaging and greedy combination strategies expanded the search range and prevented the algorithm from falling into local optima. A 3 kW fuel cell system was built to verify the correctness and effectiveness of the proposed control strategy. The experimental results showed that the proposed strategy achieved excellent control performance in terms of parameter decoupling and setpoint tracking. The strategy significantly reduced the overshoot，settling time，and oscillations，improving the response speed and dynamic performance of fuel cell system.

Optimal power allocation for electrochemical energy storage power stations based on MOIBKA algorithm

WANG Cheng, SHAO Chong, HE Xin, DONG Haiying

2025, 47(6): 74-84. doi:10.3969/j.issn.2097-0706.2025.06.008

Asbtract ( 53 )

HTML ( 0)

PDF (2199KB) ( 32 )

Figures and Tables | References | Related Articles | Metrics

To address the power allocation issue of electrochemical energy storage stations under the influence of multiple factors，an optimal power allocation strategy for electrochemical energy storage power stations based on the multi-objective improved black-winged kite algorithm （MOIBKA） is proposed. The topology of the electrochemical energy storage power station is established，and three evaluation indicators for the operation of the power station are proposed. Based on the traditional power allocation model for energy storage stations，a multi-objective power allocation model is established，aiming to achieve the lowest total operating cost of energy storage power stations，the smallest state of health loss of energy storage units，and the best state of charge （SOC） consistency. The model is solved using the MOIBKA through multiple strategies. Comparative simulation analysis and operational evaluation indicators prove that the proposed strategy could effectively reduce the number of charging and discharging cycles and the state of health loss of energy storage units while improving SOC consistency. This achieves optimal power allocation for energy storage power stations.

Residential photovoltaic power generation prediction model based on PSO-BP neural network

BAN Fengchun, CHEN Xiaofeng, HUANG Zhijia

2025, 47(6): 85-93. doi:10.3969/j.issn.2097-0706.2025.06.009

Asbtract ( 50 )

HTML ( 0)

PDF (1962KB) ( 23 )

Figures and Tables | References | Related Articles | Metrics

The residential rooftop，as an idle space that is almost free from shading，provides ideal conditions for the deployment of photovoltaic （PV） systems. However，the intermittency and volatility of photovoltaic generation，along with the mismatch between photovoltaic power generation and residential electricity demand at different times，present significant challenges for the energy management system in achieving supply-demand balance. As an essential component of energy system optimization and performance enhancement，photovoltaic generation forecasting is critical to effectively addressing these challenges. To this end，this paper proposes an improved forecasting model that combines Particle Swarm Optimization （PSO） with Backpropagation （BP） Neural Networks. In this model，PSO is employed to optimize the parameters of the BP neural network，significantly improving the accuracy and stability of photovoltaic power prediction. Experimental results demonstrate that the improved model outperforms the traditional BP neural network in forecasting accuracy across all seasons. The average root mean square error （RMSE） is reduced by 42.31%，and the coefficient of determination （R²） increases by 2.22%. The annual average forecasting accuracy exceeds 90.00%，with the highest accuracy achieved in winter，reaching 99.46%. This study provides reliable forecasting data for the optimized scheduling of photovoltaic systems in residential buildings，offering substantial practical application value.

Table of Content