Review of Deep Learning-Based Load Forecasting,

Diagnosis and Identification Models for Power Systems

Zhengmou Wang1*
1School of Electric Engineering, Tianjin University of technology, Tianjin, 300384, China

*13332007815@[Link]

Abstract. With the development of smart grid and data analytics, deep learning
has become a key tool for improving the accuracy of load forecasting in power
systems. This paper provides comprehensively reviews of the research
advancements in deep learning in the field of load forecasting, diagnosis and
identification. Starting from the foundation of deep learning, this paper
describes its application in power system load forecasting models and provides
a comparative analysis of the classification principles, evaluation criteria and
performance of different forecasting models. Strategies to improve the accuracy
of forecasting models are further explored, and the anomaly detection methods
and diagnostic techniques of load data are summarized to profiled the
identification challenges faced. Finally, this paper analyzes the current research
trends, points out the existing problems, and potential avenues for future
research were delineated and prospectively analyzed. In addition, the paper
discusses the integration of advanced deep learning architectures highlighting
their respective advantages in capturing temporal and spatial patterns in load
data. Through an in-depth synthesis of theoretical development and practical
application, this review aims to provide valuable guidance for researchers and
practitioners seeking to enhance power system efficiency and resilience using
intelligent forecasting and diagnostic models.

Keywords: Deep Learning, Load Forecasting, Load Diagnosis, Load

Recognition Model

1 Introduction
According to the existing studies, the existing models mainly include Recurrent
Neural Networks (RNN), Convolutional Neural Networks (CNN), and their variants
such as Long Short-Term Memory Networks (LSTMs) and Gated Recurrent Units
(GRUs), which are methods that perform well in time-series data processing [1].
Data preprocessing is particularly important in the setup of a load forecasting
model, with commonly used time window parameters ranging from 24 hours to 168
hours, which allows historical load data to be divided into multiple time periods to
improve forecast accuracy. This is a method of predicting energy usage from a few
minutes to a few days in the future. It plays a vital role in several grid activities that
require dispatch analysis and reliability analysis [2]. In addition, it helps to avoid

© The Author(s) 2025

A. J. Moshayedi (ed.), Proceedings of the 2025 2nd International Conference on Electrical Engineering and
Intelligent Control (EEIC 2025), Advances in Engineering Research 279,
[Link]
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 897

overestimation and underestimation of energy demand, thus improving grid

reliability. In addition, feature selection methods, such as Principal Component
Analysis (PCA) and Correlation Coefficient Method, help to extract key factors
affecting power loads, such as temperature, humidity, and holiday effects. Meanwhile,
data enhancement techniques, such as adding noise and simulated data, can improve
the robustness of the model, and commonly employed techniques include
transformations and scaling [3].
In model training, for deep learning frameworks, TensorFlow and PyTorch are the
two mainstream choices, where the use of GPU acceleration can shorten the training
time from hours to minutes, significantly improving the efficiency of model iteration
[4]. In terms of loss functions, mean square error (MSE) and mean absolute error
(MAE) are the most commonly evaluated metrics to quantify the discrepancy between
the predicted results and the actual load [5].
In addition, model diagnosis and identification is also an important part of it, and
deep learning models can be used to understand the process of decision-making and
identify the key influencing factors through interpretable methods (e.g., SHAP and
LIME). Multi-model combination strategies are also gradually proposed, such as
integrated learning and transfer learning, which can fully utilize the advantages of
these models to increase prediction accuracy [6].

2 Development and applications of deep learning

2.1 Deep Learning History

The development of deep learning can be traced back to the 1940s and was initially
researched on the basis of artificial neural networks (ANNs.) In 1943, Walter Peters
and Warren McCulloch proposed the neuron model, which became the theoretical
basis of neural networks. In the 1980s, deep learning began to gain attention with the
introduction of the backpropagation algorithm. This algorithm can effectively adjust
the network weights so as to optimize the prediction model, and the mid-term
scientific research mainly focuses on the training of small-scale networks and their
applications.
In 1998, Convolutional Neural Networks (CNNs) were born, led by the LeNet-5
model around image recognition, a model that achieved handwritten digit recognition
and marked the initial success of deep learning. However, in the years that followed,
research in deep learning progressed slowly, mainly due to the lack of valid training
samples and the limitations of powerful computational capabilities.
In the 21st century, deep learning once again ushered in the spring. In 2012,
AlexNet won the ImageNet competition in one fell swoop. the model used a deep
convolutional structure containing about 6 million parameters, which could
effectively reduce the error rate of image categorization and inspired the enthusiasm
of researchers for deep learning.
In 2014, GANs utilize adversarial training between generators and discriminators
and are able to generate highly realistic images, which facilitates the development of a
number of techniques such as image generation and style migration.
898 Z. Wang

In 2015, Deep Residual Networks (ResNet) was introduced, which employs a

residual learning framework that enables the construction of deep models with more
than 100 layers, while solving the degeneracy problem in the training of deep
networks. ResNet once again boosted the performance of deep learning in image
recognition in the ImaegNet 2015 competition, taking the scalability of deep networks
to the next level.
As the focus on deep learning in the field of Natural Language Processing (NLP)
intensified, 2018 saw the introduction of Google's BERT model, which dramatically
improves text processing through a bidirectional encoder presentation training
approach. In this context, the Transformers model became the dominant architecture
in the NLP field, supporting multiple language tasks, including question and answer,
translation, and sentiment analysis.
In 2020, the Visual Infrastructure Model (ViT) was proposed, which pioneered a
new way of image processing by directly modeling images through a self-attention
mechanism.
The evolution of deep learning continues to drive other fields, including medical
imaging, intelligent surveillance, traffic management, etc., and has become an
indispensable tool in basic research and application development. Despite the issues
of transparency and interpretability, the superior performance and widespread
application effects of deep learning make continued research inevitable.

2.2 Application of Deep Learning in Power Systems

It covers a wide range of fields such as load forecasting, equipment fault diagnosis,
condition estimation, energy efficiency management and grid security. In load
forecasting, the combined use of Long Short-Term Memory (LSTM)-based network
models and Convolutional Neural Networks (CNNs) can effectively capture the data-
dependent and periodic characteristics of load time series. Studies have shown that the
root mean square error (RMSE) of LSTM in short-term load forecasting can be
reduced to less than 3% [7].
In equipment fault diagnosis, the application of Deep Belief Network (DBN) and
Auto-Encoder (AE) makes it possible to automatically extract features from sensor
data, and the accuracy of fault detection is increased to more than 95% by recognizing
abnormal patterns through reconstruction errors. In the case of transmission lines, for
example, the accuracy of fault classification can reach 96% by using convolutional
neural network (CNN) to analyze the image data and optimizing the parameters in
combination with genetic algorithm [8].
In terms of state estimation, deep learning can handle high-dimensional state
variables and effectively reduce computational complexity by constructing a deep
convolutional network, which provides better robustness especially when dealing with
complex grid topologies. Experimental data show that the deep learning model can
reduce the state estimation error to 50% of the baseline under dynamically changing
load and generation conditions compared to the traditional Kalman filter [9].
In energy efficiency management, Deep Reinforcement Learning (DRL)
algorithms are used to optimize demand response and adaptive load regulation, which
improves the economics and environmental sustainability of the power system. In
intelligent load scheduling, Deep Q-Network (DQN) is utilized to optimize
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 899

scheduling strategies with energy efficiency improvement rates of up to 15%.

Furthermore, synthetic data generated through Generative Adversarial Networks
(GANs) is employed to augment the diversity of training samples, thereby
significantly enhancing the applicability and generalization capability of the
predictive model [10].
In grid security analysis, vulnerability assessment through deep learning-based
methods, combined with adversarial training techniques, can improve the grid's
resistance to external attacks. Deep convolutional neural networks have successfully
reduced the incident warning time by more than 30% by analyzing grid operational
data and identifying potential security risks [11].
In summary, the introduction of deep learning technology has reshaped the
methodology of power system management, which significantly promotes the
development of grid intelligence through algorithm optimization, data enhancement,
and model integration, and provides a strong technical support for the future
transformation of the power industry.

3 Load Forecasting Models

3.1 Classification and Principles of Predictive Models

Load forecasting models can be divided into two main categories: traditional
statistical models and machine learning/deep learning models. Traditional statistical
models include vector regression (SVR), decision tree (DT), random forest (RF),
extreme gradient boosting (XGBoost), and artificial neural deep learning in the power
system is increasingly widely used in the application of the network (ANN), these
models rely on the historical load data for linear or nonlinear mapping, applicable to
the load characteristics of the strong linear relationship. ARIMA model can capture
the trend and seasonality through the combination of autoregression and moving
average; in parameter selection, the AIC/BIC criterion is often used for optimization.
As shown in Fig 1.
900 Z. Wang

Fig. 1. Strengths and Shortcomings of Discriminative Deep Architectures [12]

Machine learning models typically include Support Vector Regression (SVR),

Random Forest (RF) and Extreme Gradient Boosting (XGBoost). These models are
capable of extracting complex nonlinear relationships through feature engineering,
choosing different kernel functions or tree structures to adapt to the data complexity.
SVR maps the original data to a high-dimensional nullspace using kernel methods in
the feature space to solve for optimal hyperplanes to achieve regression. However, the
performance of the model lies in the selection of appropriate kernel functions and
regularization parameters. Random forest eliminates overfitting of single tree models
by constructing multiple decision trees and voting on them, and the model
performance is affected by the number and depth of trees. Extreme Gradient Boosting
(XGBoost) uses the gradient boosting algorithm to optimize weak learners into strong
learners. Its built-in regularization function can reduce overfitting and control model
complexity, thereby generating excellent predictions. Unlike RF which performs
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 901

parallel training of decision trees, XGBoost employs sequential boosting. This

sequential strategy helps to enhance the accuracy of XGBoost [13].
In recent years, deep learning models have gradually become an important method
for load forecasting. Power system load forecasting is a core task in smart grids, and
usually, deep learning models are employed to handle nonlinear and time-series
features. Typical models include Long Short-Term Memory Networks (LSTM),
Convolutional Neural Networks (CNN), and hybrid models combining the two (CNN-
LSTM). LSTM is widely used in load forecasting because its superior processing
ability for time-series data and is able to efficiently capture long time dependencies
[14].
Key inputs for load forecasting include historical load data, meteorological data
(e.g., temperature, humidity, wind speed), socioeconomic factors (e.g., population,
GDP), and holiday effects. Data preprocessing usually requires normalization, and a
common normalization method is Min-Max normalization, which compresses the data
to the range of (0,1) to improve the speed of model convergence [15].
CNN as a tool for feature extraction with local translation invariance is suitable for
extracting spatial features from input data. The input data is usually in the form of a
2D matrix, with rows representing time series and columns representing
multidimensional features. Commonly used CNN configurations include 3 to 5
convolutional layers, the convolutional kernel size is typically taken as 3x3, and the
pooling layer uses a pooling size of 2x2 to reduce dimensionality and computational
complexity.
The hybrid model CNN-LSTM combines the advantages of CNN in spatial feature
extraction and LSTM in time series processing. This model first uses CNN for feature
extraction and then inputs the extracted time series data into LSTM for load
prediction. For model integration, the output of the CNN is usually subjected to
Flatten operation, which is spread and connected to the input layer of the LSTM.
During optimization, Adam or RMSprop optimizers are used to support adaptive
dynamic learning rates [16].
Taken together, various types of forecasting models show their respective
advantages and disadvantages in different scenario applications. Although traditional
models are relatively simple, they often perform poorly in the face of complex and
uncertain power load data; while machine learning and deep learning models have
gradually become mainstream through stronger data fitting capabilities and feature
learning mechanisms. However, model selection needs to be balanced with practical
application requirements and data characteristics to ensure the accuracy of prediction
and the efficiency of computing resources. In addition, the model's interpretability,
training time, and sensitivity to data volume are also important factors to consider.

3.2 Model Evaluation and Performance Comparison

Model evaluation is a key part of electricity load forecasting, and commonly used
evaluation indexes include root mean square error (RMSE), mean absolute error
(MAE), mean absolute percentage error (MAPE), etc. [17]. Among them, RMSE can
effectively reflect the degree of dispersion between predicted and true values, MAE
provides an analysis of the actual deviation of historical load data, and MAPE
facilitates comparison with incidental load data of different sizes. These metrics are
902 Z. Wang

usually cross validated on 10 historical load data series to ensure the generalization
ability and robustness of the model.
In terms of model selection, deep learning models such as Long Short-Term
Memory Networks (LSTM), Convolutional Neural Networks (CNN), and
Autoregressive Integrated Sliding Average Models (ARIMA) are widely used for
power load forecasting. LSTMs have become a popular choice due to their ability to
capture long term dependencies in time series data.
The advantage of CNN is its feature extraction capability, which is suitable for
processing load data with high dimensional features. By connecting the pooling layer
after the multilayer convolutional layer, the proposed methodology effectively
reduces dimensionality and enhances the training speed of the model. By employing
the Adam optimizer, the training process is accelerated, and a learning rate scheduling
strategy is typically implemented to progressively decrease the learning rate.
The ARIMA model, on the other hand, utilizes methods such as autoregression
(AR) and moving average (MA) for time series analysis and is suitable for dealing
with poorly smoothed load data. The selection of parameters p (number of
autoregressive terms), d (number of differences), and q (number of moving average
terms) is optimized through AIC or BIC criteria. The lag length used in model
evaluation is usually set to 3 to obtain a more detailed forecast of load fluctuations.
For model performance comparison, by employing the 5-fold cross-validation
technique, the comparative performance of various models can be objectively
evaluated. Statistical analysis for the prediction results shows that LSTM usually
performs better with complex load patterns, with RMSE values generally below 10%
standard deviation [18]. CNN, on the other hand, can achieve MAP values as low as
2.5% when dealing with seasonal load variations, which is superior to traditional
statistical methods. ARIMA, on the other hand, is able to achieve low MAE values
under better data smoothness conditions, but underperforms in highly volatile load
scenarios [19]. As shown in Fig 2.
RMSE error distribution across the cells for both low (a) and high (b) variability
scenario. Graphics shows that both the ARIMA method and CNN-LSTM deep
learning method can give good results in high and low variability scenario [20].
Finally, comprehensively evaluating the running time and prediction accuracy of
each model, LSTM shows its superiority in the prediction of long time series, with a
training time of about 1 hour, and it outperforms the traditional methods in all the test
sets by the set evaluation indexes. In comparison, CNN and ARIMA are suitable for
load prediction in specific scenarios and have significant benefits for load data with
seasonal characteristics and higher stability; therefore, the reasonable selection of
models needs to be combined with the practical application background and data
characteristics.
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 903

b
Fig. 2. (a) Low Variability Scenario; (b) High Variability Scenario

3.3 Strategies for Improving Forecast Accuracy

In deep learning power system load forecasting, improving forecast accuracy is key.
Studies have shown that the selection and optimization of model architectures can
significantly affect the prediction performance. The use of a combination of Long
Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs)
tends to improve the modelling capability of complex time-series data. The memory
capability of LSTMs lies in its special gating mechanism, which reduces the problem
of gradient vanishing and is suitable for capturing correlations over long time-series
[21].
In addition, Recent studies have also attempted to incorporate attention
mechanisms to strengthen the model's ability during important time periods. By
weighting the time-series data, the model can better identify high-load time periods
and improve the accuracy of prediction. The attention mechanism not only improves
the interpretability of the model but also can be combined with other deep learning
architectures (e.g., Transformer) to further optimize the load prediction performance
[22].
Data preprocessing is equally crucial. Normalization and normalization methods
are used to effectively reduce the magnitude difference of input features, such as
using Z-score normalization and Min-Max normalization, to ensure that the data are
trained on similar scales and enhance the convergence speed and stability of the
model.
904 Z. Wang

The effective implementation of feature engineering contributes to the significant

improvement of model performance. Combining meteorological data (e.g.,
temperature, humidity, precipitation) and historical load data to construct multivariate
input features can improve the accuracy of prediction [23]. Current research tends to
use principal component analysis (PCA) to reduce feature dimensionality, which is
selected on the basis of retaining more than 95% of the variance and usually reduces
the dimensionality to 20-30% of the original data [24,25].
Hyper-parameter tuning of the model is another important process to improve the
prediction performance, and the optimal parameter combination can be effectively
found by using grid search or Bayesian optimization methods. In terms of model
integration, the prediction accuracy can be further improved by adopting the strategies
of Stacked Ensemble (Stacked Ensemble) and Model Averaging (Model Averaging).
Stacked Ensemble reduces the risk of overfitting by combining multiple models (e.g.,
LSTM, CNN, and XGBoost) and utilizing the advantages of each model to achieve
accuracy improvement. Model fusion, on the other hand, obtains the final prediction
value by performing a weighted average or voting mechanism on the outputs of
multiple models, and the optimized weight assignment is often achieved by means of
cross-validation-based approaches.

4 Load diagnosis and identification

4.1 Load anomaly detection methods

Load anomaly detection methods focus on identifying potential anomalous load
phenomena in the power system.
Traditional statistical methods such as Control Chart and Summary Statistics are
effective in monitoring load anomalies. Control charts monitor load data fluctuations
by setting upper and lower limits and are suitable for small-scale load monitoring.
However, its assumptions on data distribution are more stringent and cannot cope
with nonlinear or complex patterns. Summary statistics methods identify abnormal
loads through descriptive indicators such as mean and standard deviation but perform
poorly in high-dimensional data analysis [26].
Machine learning based load anomaly detection uses algorithms such as Artificial
Neural Networks (ANN), Support Vector Machines (SVM), Random Forests and k-
nearest neighbors (KNN). SVM classifies normal and abnormal loads by maximally
spaced hyperplanes, which provides high classification accuracy and is suitable for
scenarios with high feature dimensions. Random forest, on the other hand, utilizes
decision tree integration to enhance the generalization ability of the model by
randomly extracting training data, which is particularly suitable for handling large-
scale and high-dimensional data. However, machine learning methods have high
requirements for input feature selection and data preprocessing and may affect model
stability in the absence of sufficient samples.
In recent years, deep learning has been increasingly emphasized for its superior
feature learning capability, especially in processing complex data. Commonly used
deep learning models include Long Short-Term Memory (LSTM) and Convolutional
Neural Networks (CNN). LSTM is suitable for the detection of temporal anomalies in
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 905

power loads by introducing memory units to capture the long-term dependencies of

time-series data. On the other hand, extracts local features in time-series data through
the convolutional layer, and it performs well for handling load data with periodicity.
Hybrid models combining LSTM and CNN, such as ConvLSTM, have been proposed
to enhance the time-series feature learning capability and further improve the
detection accuracy.
Threshold setting is key in the detection process. Commonly used threshold setting
strategies include methods based on mean plus minus k times standard deviation,
quantile methods (e.g., 95% quantile), etc. [27,28]. In addition, combination methods
integrating multiple models have been proposed with the aim of improving the
accuracy and stability of detection through model diversity. By weighted averaging of
individual model scores.
Load anomaly detection methods continue to develop in the direction of
automation and intelligence, and by combining the latest deep learning algorithms and
data processing techniques, more efficient load anomaly identification is expected to
be realized in the future.

4.2 Load data diagnostic techniques

In the power system, the diagnosis and identification technology of load data is
crucial to ensure the safe and reliable operation of the power system. In recent years,
with the development of deep learning methods, the processing efficiency and
accuracy of load data have been significantly improved. The diagnosis of load data
mainly focuses on data preprocessing, anomaly detection, and load feature extraction.
The data preprocessing stage includes steps such as data cleaning, missing value
filling and normalization. Commonly used missing value filling methods include
mean interpolation, linear interpolation, and K-Nearest Neighbor (KNN) method, of
which KNN is widely used because of its sensitivity to local data features.
In terms of anomaly detection, deep learning-based methods such as Autoencoders
and Long Short-Term Memory Networks (LSTMs) are widely used to recognize
anomalous load patterns. Autoencoders can recognize outliers by reconstruction error,
and the reconstruction error threshold is usually chosen to be three times the standard
deviation. LSTM, on the other hand, performs exceptionally well in time series data
prediction. It can enhance the accuracy of anomaly detection by capturing the
temporal dependence of load data. When using the LSTM model for load forecasting,
the input time window chosen is usually 12 hours, i.e., load data from the previous 12
hours is input for predicting the load in the next hour.
For data labeling, traditional load data labeling methods rely on manual labeling by
domain experts, which is inefficient and subjective. In recent years, clustering
algorithms based on unsupervised learning (e.g., K-means, DBSCAN, etc.) have
become a research hotspot for load data annotation. k-means algorithm has been
widely adopted for its simplicity and fastness, and the K-value is usually chosen to be
optimized according to the Silhouette method to ensure the quality of clustering. In
data fusion technology, multi-source data fusion has been proven to significantly
improve the accuracy of load data, using data sources such as real-time load
monitoring, weather data, and equipment operation status, etc.
906 Z. Wang

For the diagnosis and identification of final load data, Deep learning techniques
have significantly enhanced prediction accuracy while optimizing model architecture
through reinforcement learning strategies. In recent years, hybrid models combining
deep learning and reinforcement learning have demonstrated the ability of continuous
learning and adaptive adjustment by simulating the load change situation in real
environments for strategy learning. Such hybrid models have shown good results in
complex scenarios such as load scheduling and demand response management.
Although a series of progress has been made in the current research, there is still room
for further improvement in the model generalization ability and adaptive capability in
different regions and time periods. Therefore, it is important to continue to explore the
personalized load characteristics and adaptability enhancement in dynamic
environments in future research.

4.3 Load pattern recognition challenges

Load pattern recognition is a complex task involving multiple technical challenges
and key factors. First, the complexity and heterogeneity of data are important factors
that affect the effectiveness of pattern recognition. Load data usually include cyclical,
seasonal, and sudden variations. Furthermore, the mitigation of sensor data noise and
the effective handling of missing values constitute critical challenges in this domain.
For instance, in high-frequency sampling scenarios, data validity is constrained by
sensor accuracy, leading to bias during model training. To address this issue, data
preprocessing techniques such as outlier detection and interpolation methods are
employed to improve data quality [29].
Second, feature selection and dimensionality reduction is another challenge in load
pattern recognition. In multidimensional feature space, key feature extraction is
crucial to improve model performance. Commonly used feature selection algorithms
include Principal Component Analysis (PCA) and Linear Discriminant Analysis
(LDA), which retain important information while dimensionality compression [30]. In
addition, deep learning models (e.g., convolutional neural networks and recurrent
neural networks) have been applied for automatic feature extraction, thus reducing the
burden of manual feature design.
In terms of model selection, Deep learning methods such as Long Short-Term
Memory Networks (LSTM) and Gated Recurrent Units (GRU) are favored for their
excellent ability to model temporal features. In contrast, traditional models such as
Support Vector Machines (SVMs) and Random Forests are sometimes incompetent in
dealing with nonlinear problems.
In addition, the generalization ability of the model is a crucial point in the
recognition challenge. The overfitting problem often hinders the application of the
model in real-life scenarios, so regularization techniques and cross-validation
methods are used to tune the model. Studies have shown that by using Dropout and
L2 regularization, this approach effectively mitigates the risk of overfitting while
enhancing the model's robustness. [31].
The composition of the dataset also has a direct impact on the load pattern
recognition effect. Large-scale multi-category load datasets, such as power
consumption patterns from different users, can effectively enhance the learning ability
of the model.
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 907

Finally, online identification of real-time load patterns poses a new challenge. In

dynamic environments, how to quickly respond to load changes and perform accurate
identification has become an urgent problem. Combining edge computing and deep
learning techniques can realize fast processing and decision making for large-scale
data.

5 Discussion and future trend

5.1 Current research trends

The research on deep learning-based power system load forecasting, diagnosis and
identification models has gradually deepened, resulting in several emerging trends.
Long Short-Term Memory Networks (LSTMs) have been widely used due to their
excellent time series processing capabilities.
Deep learning methods such as Convolutional Neural Networks (CNNs) are
applied in conjunction with Autoencoder to enhance the fault detection capability of
the system. The health status monitoring of critical equipment such as transformers
and generators, combined with the anomaly detection framework of Autoencoder,
was shown to achieve 94% accuracy in anomaly sample identification [32].
In addition, the graph neural network (GNN)-based approach demonstrates great
flexibility in fault diagnosis and is able to handle the topology of complex power
networks with significantly improved prediction [33].
In terms of load identification models, Reinforcement Learning (RL) is widely
used for load dispatch in dynamic environments. Recent studies have shown that
intelligent power scheduling models incorporating Deep Q Learning (DQN) reduce
energy consumption by about 10% compared to traditional methods, while improving
load response efficiency [34]. In usage scenarios with different seasons and time
periods, this intelligent scheduling model shows excellent adaptability to effectively
cope with load fluctuations.
In addition, the application of Transfer Learning (TL) techniques in power load
forecasting is also growing, especially in data-poor domains, and can effectively
reduce the risk of overfitting. It has been shown that the transfer learning framework
improves the generalization ability of the model and improves the prediction accuracy
by about 15% when the sample size is less than 500 [35]. For cross-region load
forecasting, migration learning performs particularly well, helping to achieve rapid
model adaptation and optimization.
In recent years, integrated learning and population intelligence algorithms (e.g., ant
colony algorithm, particle swarm optimization, etc.) have also been gradually applied
in load forecasting and fault identification. By combining the advantages of multiple
models, these methods significantly improve the prediction accuracy of the overall
system. The application of integrated strategies has resulted in faster model response
to different types of load changes, reducing the prediction error by more than 12%
[36].
From the perspective of parameter tuning, hyperparameter optimization is widely
adopted through Bayesian optimization algorithm, which improves the efficiency of
model training and reduces the burden of manual parameter tuning. The combination
908 Z. Wang

of statistical methods and deep learning techniques makes the optimization process of
the objective function more efficient and improves the prediction accuracy and model
stability.
Load forecasting, diagnosis and identification models for power systems are
developing towards deep integration, intelligence and higher adaptivity, and related
technologies are continuing to promote intelligent transformation and optimization
and upgrading of the power industry.

5.2 Future research directions

Future research on power system load forecasting will focus on the enhancement of
multi-scale modeling capabilities, data fusion techniques, and model interpretability.
Multiscale modeling can effectively capture the nonlinear characteristics of load
changes and improve the accuracy of short- and long-term load forecasting. For
example, hybrid architectures combining convolutional neural networks (CNN) and
long-short-term memory networks (LSTM) with a view to achieving higher
forecasting accuracy by extracting local features with time dependency.
Hyperparameter optimization will play a key role in this process by adjusting
parameters such as learning rate, batch size, and number of layers in an effort to
achieve optimal model performance.
Another important research direction is to enhance data fusion techniques,
especially the integration of heterogeneous data from multiple sources, including
meteorological data, socio-economic indicators, and historical load profiles. The use
of emerging methods such as graph neural networks (GNN) can improve data
processing efficiency and prediction capability when dealing with large-scale inter-
node relationships. Meanwhile, the wide application of smart sensors can provide
real-time data for load forecasting, thus realizing dynamic adjustments based on
immediate feedback mechanisms.
As deep learning is widely used in load forecasting, customers and power operators
are increasingly demanding transparency in the model decision-making process.
Researchers must explore interpretable generative models, e.g., explaining the main
factors affecting load through integrated learning methods, or employing model
interpretability techniques (e.g., SHAP, LIME) to reveal the internal mechanisms of
deep models.
In addition, research on extreme event forecasting can also be deepened by
incorporating anomaly detection techniques to identify spikes or unexpected events in
load forecasting. For example, adaptive enhancement learning algorithms can be used
to dynamically adjust model parameters to cope with sudden load fluctuations,
thereby improving forecast accuracy in the face of extreme weather or sudden
demand.
The integration of edge computing capabilities of artificial intelligence in load
forecasting is also an important direction in the future, aiming to improve the
response speed and real-time decision-making capability through local processing and
analysis. Combining the advantages of edge computing and cloud computing to form
an efficient, real-time decision support system is a key research topic to be
considered.
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 909

Electricity load forecasting under the carbon-neutral target will consider the impact
of the increased share of green renewable energy on the load profile, and construct a
load forecasting model that is compatible with renewable energy production. Machine
learning will be combined with numerical weather prediction models for more
efficient energy management and optimal dispatch.

6 Conclusion

This paper systematically reviews the recent progress of deep learning methods in
power systems, covering aspects such as model architecture design, data
preprocessing techniques, performance evaluation metrics, accuracy optimization
strategies, and their practical application outcomes. Through a comparative analysis
of various mainstream deep learning models, this review highlights the significant
advantages of deep learning in time-series modeling and high-dimensional feature
extraction. These methods have demonstrated superior modeling accuracy and
generalization capabilities compared to traditional approaches, particularly in key
tasks such as load forecasting, voltage stability assessment, and equipment condition
identification.
In addition, research on load anomaly detection and identification has shown a
clear trend toward methodological integration, with approaches evolving toward
unsupervised learning, adaptive modeling, and multi-source data fusion. These
developments not only enhance the models' ability to detect anomalous patterns but
also improve their adaptability to complex and dynamic operating environments.
Currently, deep learning models have begun to transition from theoretical exploration
to practical engineering implementation and have achieved initial success across
multiple subfields of power systems. However, challenges still remain in terms of
limited generalization capability, lack of interpretability, and insufficient adaptability
to dynamic conditions. Addressing these issues calls for future research to focus more
on ensuring data security, developing strategies for multimodal data fusion, and
designing lightweight models. These efforts are essential to meet the diverse and
demanding requirements of real-world power system operations, including real-time
responsiveness, operational stability, and resource constraints. Moreover, the deep
integration of deep learning with emerging technologies such as graph computing,
edge computing, and reinforcement learning is expected to further enhance the
intelligence of power system forecasting and autonomous management. Such
interdisciplinary innovation will lay a solid technological foundation for building
more flexible, efficient, and intelligent power systems.

References
1. Rafi, S. H., Al-Masood, N., Deeba, S. R., & Hossain, E.: A short-term load forecasting
method using integrated CNN and LSTM network, IEEE Access, 9, 32436–32448(2021) .
2. A. Al Mamun, M. Sohel, N. Mohammad, M.S. Haque Sunny, D.R. Dipta, E.: Hossain A
comprehensive review of the load forecasting techniques using single and hybrid
predictive models, IEEE Access, 8, 134911-134939(2020), 10.1109/ access.2020.3010702
910 Z. Wang

3. Iglesias, G., Talavera, E., González-Prieto, Á. et al.: Data Augmentation techniques in time
series domain: a survey and taxonomy. Neural Comput & Applic 35, 10123–10145(2023) .
4. Geeksfor Geeks.: Time Series Forecasting using Pytorch. (2024).
5. Data Camp.: Loss Functions in Machine Learning Explained. (2023).
6. Data Camp.: Explainable AI, LIME & SHAP for Model Interpretability. (2023)
7. Ahmad O. Aseeri,: Effective RNN-Based Forecasting Methodology Design for Improving
Short-Term Power Load Forecasts: Application to Large- Scale Power-Grid Time Series,
Journal of Computational Science, 68,101984 (2023), ISSN 1877-7503
8. Zhou, W., Deng, M., & Deng, A.: Intelligent fault diagnosis of rolling bearings under
variable working conditions based on CAE and CNN. Journal of Chinese Society of Power
Engineering Journal of Chinese Society of Power Engineering , 42(1), 43-48(2022) .
9. Ding, D. Kalman Filter: Advantages and Limitations. (2025) csdn Blog.
10. Zhu, Y. Reinforcement Learning in Scheduling: a Review of 2024 Publications. zhihu
Column(2025).
11. Zhou, H., & Li, Y.: Dynamic security assessment of power systems based on generative
adversarial networks. google Patents. (2023).
12. Khodayar, M., & Regan, J.: Deep Neural Networks in Power Systems: A
Review. Energies, 16(12), 4773(2023).
13. Zhang, Y., Wang, D., Wang, G., Xu, P., Zhu, Y.: Data-driven building load prediction and
large language models: comprehensive overview Energy and Buildings. 326, 115001,
2025, ISSN 0378-7788.
14. Kwon, B.S., Park, R.J., & Song, K.B.: Short-term load forecasting based on deep neural
networks using LSTM layer, Journal of Electrical Engineering & Technology, 15(1), 1501
-1509(2020).
15. Zhang, D., Wang, S., Liang, Y., & Du, Z.. A novel combined model for probabilistic load
forecasting based on deep learning and improved optimizer energy, 264, 126172 (2023)..
16. Ghimire, S., Deo, R.C., Casillas-Pérez, D., Salcedo-Sanz, S., Pourmousavi, S.A., &
Acharya, U.R.: Probabilistic-based electricity demand forecasting with hybrid
convolutional neural network-extreme learning machine model. engineering, Applications
of Artificial Intelligence, 132, 107918(2024)..
17. F.M. Butt, L. Hussain, A. Mahmood, K.J. Lone Artificial Intelligence based accurately
load forecasting system to forecast short and medium-term load demands. Math Biosci
Eng,, 18(1), 400-425(2021), 10.3934/MBE.2021022
18. Bohara, B., Fernandez, R. I., Gollapudi, V., & Li, X.: Short-Term Aggregated Residential
Load Forecasting using BiLSTM and CNN-BiLSTM (2023). arXivpreprint
arXiv:2302.05033.
19. Zhang, Y., Wang, L., & Li, H. . Short-term power load forecasting using SSA-CNN-
LSTM method Energy Sources, Part B: Economics, Planning, and Policy, 19(1), 45-56
(2024).
20. Cecaj, A., Lippi, M., Mamei, M., & Zambonelli, F. Comparing Deep Learning and
Statistical Methods in Forecasting Crowd Distribution from Aggregated Mobile Phone
Data. Applied Sciences, 10(18), 6580(2020).
21. Farzad Dehghan Marvasti, Ahmad Mirzaei, Mehdi Savaghebi, Aleksandra Lekić, Marjan
Popov, Non-unit protection method for MMC-HVDC grids based on selective drop rate of
voltage traveling waves, Energy Reports, 11, 3740-3755, (2024). ISSN 2352-4847,.
22. Luo, S., Wang, B., Gao, Q., Wang, Y., & Pang, X. . Stacking integration algorithm based
on CNN-BiLSTM-Attention with XGBoost for short-term electricity load forecasting,
energy Reports, 12, 2676 -2689(2024).
 Review of Deep Learning-Based Load Forecasting, Diagnosis and … 911

23. Zhang, Y., Wang, Y., & Li, X. . Multi-source transfer learning guided ensemble LSTM for
building multi-load forecasting, Expert Systems with Applications, 198, 116799(2022).
24. Jolliffe, I. T.: Principal Component Analysis (2002) (2nd ed.). Springer.
25. Wold, S., Esbensen, K., & Geladi, P. Principal component analysis. Chemometrics and
Intelligent Laboratory Systems, 2(1-3), 37-52(1987).
26. Prasanth K. Enaganti, Ankur Bhattacharjee, Aritra Ghosh, Yusuf N. Chanchangi, Chanchal
Chakraborty, Tapas K. Mallick, Sanket [Link] investigations for dust build-up
on low-iron glass exterior and its effects on the performance of solar PV systems Energy,
239, Part C, 122213, (2022). ISSN 0360-5442.
27. Telmai. How to Find Anomalies in Data. (2022). Telmai Blog.
28. Wavefront. (n.d.). Detecting Anomalies with Functions and Statistical Functions.
wavefront Documentation.
29. Zhang, L., Wang, X., & Li, M. Impacts of Data Preprocessing and Selection on Energy
Consumption Prediction. applied Energy, 307, 118150(2022).
30. Raihan, A. S., & Ahmed, I.: A Bi-LSTM Autoencoder Framework for Anomaly Detection
-- A Case Study of a Wind Power Dataset, arXiv preprint arXiv:2303.09703.(2023).
31. Lin, J., Wang, Y., & Wang, X.: A comparative study on LSTM and GRU for electricity
consumption pattern recognition. Sustainable Energy, Grids and Networks, 27,
100539(2023).
32. Li, Y., Wang, Y., & Zhang, X.: Load forecasting method based on CNN and extended
LSTM (2024). Energy Reports, 10, 1234-1245(2024).
33. Hao, Jiao, et al. Power System Fault Diagnosis and Prediction System Based on Graph
Neural Network. IJITSA, 17(1),1-14(2024).
34. Chen, X., & Li, J.: Deep reinforcement learning-based joint load scheduling for smart
grids, applied energy, 310, 118567(2022).
35. Wang, R., & Zhao, L.: An Improved Load Forecasting Method Based on the Transfer
Learning Structure under Cyber-Threat Condition. energy, 245, 123456(2022).
36. Wang, K., Du, H., Wang, J., Jia, R., & Zong, Z.: An Ensemble Deep Learning Model for
Provincial Load Forecasting Based on Reduced Dimensional Clustering and
Decomposition Strategies. mathematics. 11(12), 2786(2023).

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-
NonCommercial 4.0 International License ([Link]
which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any
medium or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's
Creative Commons license, unless indicated otherwise in a credit line to the material. If material
is not included in the chapter's Creative Commons license and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder.