Qiu et al. ^[11][75] propose a hybrid incremental learning approach that comprised discrete wavelet transform, empirical mode decomposition, and random vector functional link network. By using the ensemble method, the efficiency and accuracy of STLF can be improved. Li et al. ^[12][76] present a model with an ensemble approach that integrates three base methods for STLF in which the experiments show the model’s effectiveness for STLF. However, the choice of base methods in the ensemble approach needs further validation. Many DL-based methods are used to solve LF problems. In recent years, DNNs have been used to obtain the potential knowledge for a forecasting model. However, the ANN method is often trapped in local minima ^[13][77] and over-fitting problems. Shi et al. ^[14][78] proposed a pooling-based deep RNN for STLF to address the over-fitting issue by increasing data diversity and volume. To address the time-consuming procedure of building a optimal DNN, which determines the number of hidden layers in the DNN model, Moon et al. ^[15][67] used an ensemble method that combines multiple DNN models with different numbers of hidden layers to achieve overall better performance by eliminating the poorly performed models. However, the computing overhead is a limitation, because several CNNs are included. In He, Deng, and Li ^[16][79], a DBN embedded with parametric Copula models, is proposed to forecast the hourly load of a power grid of an urban area in Texas, and the results reflect the effectiveness of the method by comparing it with neural networks, SVR, and ELM. Hafeez et al. ^[17][43] propose a hybrid algorithm using factored conditional restricted Boltzmann machine (FCRBM) as a training module and genetic wind-driven (GWDO) as an optimization algorithm. The model is validated by outperforming the state-of-the-art algorithm. Aly ^[18][80] built a hybrid clustering method based on wavelet neural network (WNN) and ANN schemes and showed the higher performance of the proposed model, comparing it with other clustering methods.

Even though the majority of LF problems fall into STLF, MTLF and LTLF are also very crucial for stable and smooth power system operation. MTLF is used to coordinate load dispatch, maintenance scheduling, and balance demand and generation ^[19][81]. Unlike STLF, which fit data to a model, MTLF and LTLF have different problems that are often ignored due to their complications ^[20][82] and randomness ^[21][83]. The MTLF and LTLF are not only affected by some explicit factors, such as historical load and weather data, but are also affected by local economy and demographic data, such as population and appliances in use ^[19][81]. Unlike STLF, which treats all weather variables with equal importance, the weather indicators for MTLF and LTLF follow a decreasing order of importance from temperature, humidity, wind, and precipitation ^[22][84]. Jiang et al. ^[23][85] proposed a dynamic Bayes network (DBN)-based MTLF model to forecast the peak power load for the following year. In Askari and Keynia ^[24][86], the authors deployed a DNN model with an optimized training algorithm that comprises two search algorithms for MTLF in power systems and presented the effectiveness of the model. Liu et al. ^[25][87] also provided a neural network-based model with particle swarm optimization (PSO) and showed the feasibility and validity of the model. Rai and De ^[26][88] improved a support vector regression model for MTLF with an average minimum mean absolute percentage error (MAPE) of 3.60. Gul et al. ^[27][89] provide a solution based on CNN and LSTM methods. Dudek et al. ^[28][90] propose a hybrid DL model for MTLF that combines exponential smoothing, advanced LSTM, and the ensemble method. This is a competitive method that also uses the ensemble approach.

LTLF is used to predict the power consumption, system planning, and scheduling of generation units expansion in power systems. Generally, it spans from a few years to a couple decades. Because it needs a huge investment to construct new power generation, it requires accurate and effective forecasting for power systems. There are many ML and AI techniques developed for the problem. Nalcaci et al. ^[29][37] show that the MARS method gives more accurate and stable results than ANN and LR models when predicting the relationship between load demand and several environmental variables. Ali et al. ^[30][91] applied a novel hybrid fuzzy-neuro model for LTLF. LSTM is also well used in the domain. In 2017, Zheng et al. ^[8][72] exploited the LSTM-based RNN for the long-term dependencies in the electric load time series for LTLF, in which the method had a promising performance. Agrawal et al. ^[31][92] also propose an LTLF model with hourly granularity by using the LSTM network with high accuracy. To solve the vanishing and exploding gradient problems of LSTM, Dong et al. ^[32][93] present a hybrid method based on LSTM and gated recurrent unit (GRU) with a good performance for LTLF. In Kumar et al. ^[33][94], Apache Sparks was used to deploy a hybrid model that comprises LSTM and GRU for hyperparameter tuning purposes. Bouktif et al. ^[34][95] also proposes an LSTM-RNN model for this task. Sangrody et al. ^[35][96] compared six commonly used ML technologies: ANN, SVM, RNN, KNN, GPR, and generalized regression neural network (GRNN). ANN showed better performance than the other five methods for LTLF. Table 1 summarizes the AI techniques for LF.

Transient stability assessment (TSA) is the ability to determine whether a system will remain synchronised after a huge perturbation. The two most commonly used traditional methods for TSA are time domain simulations and direct methods. However, the increasingly complex power systems result in great challenges in making reliable decisions based on traditional TSA methods.

Fortunately, the development of AI technologies provides the new prospective methods to this issue by using the large volume of data collected by PMU and WAMS. In Baltas et al. ^[38][99], three ML algorithms—decision trees, SVMs, and ANNs, which are for online TSA—were compared by using two datasets. The results show similar performance for the methods, and performance varies according to dataset quality. Mahdi et al. ^[42][103] also used a trained ANN model for online TSA prediction with promising performance. Hu et al. ^[43][104] developed two improved SVM methods to solve the traditional SVM limitation that reduces the false and missed alarms. Mosavi et al. ^[44][105] present a deep neuro-classifier for TSA and showed the high-generalization capacity of the model. Tang et al. ^[45][106] propose a TSA method that combined trajectory fitting (TF) and ELM, and the hybrid method showed effectiveness and reliability. Yu et al. ^[46][107] propose an RNN-LSTM model that better learns from the temporal data dependencies of the input data. Tan et al. ^[47][108] built a supervised classifier that consists of CNN and stacked autoencoders (SAE) for TSA problems with high accuracy. Liu et al. ^[48][109] used an intelligent system that comprised an ensemble of neural networks based on ELMs with 100% accuracy. In 2020, the study ^[49][110] applied a deep belief network (DBN) for TSA with great accuracy improvement. Shi et al. ^[50][111] trained a CNN model to provide a solution for online TSA for power system control.

Power grid frequency stability assessments (FSAs) can be defined as the ability of a system to maintain a steady range of frequency following a severe system upset or perturbation that results in an imbalance between generation and load ^[37][98]. A large frequency deviation causes generation units to trip, and the system stability can eventually be influenced. A few studies focused on this area by using AI technologies. In 2019, Wang et al. ^[51][14] proposed a hybrid model that integrated a frequency response model with an extreme learning ML model for FSA.

Small-signal stability is defined as the ability of the system to maintain synchronism when it is under small disturbances ^[52][112]. The term “small-signal stability assessment” is interchangeable with the term “oscillatory stable assessment” (OSA). A CNN-based method ^[50][111] was also developed for OSA, and the results show that the model is robust to PMU noise and that algorithm performance will not be reduced as the system grows in scale. Xiao et al. ^[53][113] used a multivariate random forest regression (MRFR) algorithm for OSA on an 18 bus test system, and the results presented high accuracy and robustness. Kamari et al. ^[54][114] deployed a PSO scheme to accelerate the determination of OSA.

Voltage collapse can significantly influence the stability of power systems. Thus, a voltage stability assessment (VSA) model, which can evaluate the voltage stability of the system in a timely fashion, would be a prevention. Numerous AI-based models are proposed in VSA, such as ANN ^[55][115], SVM ^[56][116], decision trees ^[57][117], and FL ^[58][118]. Ashraf et al. ^[55][115] used an ANN model to estimate the loading margin of power systems and testified to the effectiveness on Institute of Electrical and Electronics Engineers 14 bus and 118 bus test systems. Amroune et al. ^[59][119] used a hybrid model by using dragonfly optimization and SVR for online VSA. Mohammadi et al. ^[56][116] proposes a method for VSA by using an SVM. The results showed that the misclassification rates of the SVMs are as low as 2% for real power grids. Yang et al. ^[60][120] built a moment-based spectrum estimation method to gain insight into changes of voltage magnitudes for real-time static VSA. In Meng et al. ^[57][117], a decision tree model was used for online VSA. Liu et al. ^[61][121] built a feature selection model using partial mutual information (PMI) on an iterated random forest (IRF) model. An in-depth review is also found in Amroune ^[62][122]. Table 2 summarizes the AI techniques for the power system stability assessment.