According to recent studies, metro systems are impacted by various weather conditions, resulting in both positive and negative effects. For example, warmer weather typically leads to increased ridership, while cold and windy conditions tend to decrease transit use [14] . The impact of weather on public transportation varies widely across different systems. A study analyzing daily metro ridership in Nanjing from 2011 to 2014 found that certain transportation modes are more resilient to adverse weather than others. It was also noted that weekend travelers tend to be more affected by weather conditions compared to weekday passengers [15] . Additionally, the influence of weather on travel behavior depends on the mode of transportation and its perceived comfort. Extreme weather affects different modes differently, with private cars often seen as more comfortable and reliable during warmer conditions, potentially reducing subway usage. However, subways generally maintain steady ridership due to commuters’ adaptability in switching between transportation modes [16] . Weather fluctuations can also impact metro operations, leading to reliability issues and increased operational costs [17] . The exacerbation of these challenges by climate change is noted, as severe weather events impact leisure travel more than daily commuting. For instance, in New York City, rain and snowfall typically decrease public transport usage, while lower-than-normal temperatures may have a positive effect [18] . Notably, based on hourly and station-level data, weather was observed to have the most significant impact on passenger volume in the afternoon, followed by midday and morning [19] . Moreover, the time of day significantly influences passenger arrival volumes, categorizing travelers into schedule-dependent and independent groups. This distinction is evident across morning, evening, and off-peak periods, affecting how passengers choose their travel times based on metro schedules [20] . The study also revealed that passenger behavior and arrival patterns vary accordingly, with peak hours typically experiencing higher volumes compared to off-peak times [21] . Therefore, to comprehensively understand and predict passenger arrival trends throughout the day, it’s essential to analyze these variations across different timeframes.

As per the findings of Jolliffe et al. [22] - [24] , the frequency of train services is a crucial factor in shaping passenger arrival patterns. The availability of gaps in public transportation significantly affects these patterns. The impact of train frequency on passenger arrivals is manifested in various ways. Higher frequencies result in less predictable arrival patterns, leading to an average waiting time of approximately half the headway. Conversely, lower frequencies prompt passengers to strategically time their arrivals to minimize waits [25] . Research has indicated that the shift from random to non-random arrival patterns typically occurs within headways ranging from 5 to 11 minutes, regardless of frequency levels. Passenger behavior also plays a significant role in shaping arrival patterns, with individuals categorized into those adhering to schedules and arriving on time, and those arriving randomly [26] .

The unpredictability and fluctuation of passenger arrivals over time can significantly impact operational efficiency, particularly during peak hours when arrival flows vary dynamically. Addressing this uncertainty often involves assuming probability distributions for arrival volumes, although accurately defining such distributions can be challenging in practical applications. Therefore, considering a range of variability in arrival volumes is often deemed more practical than relying solely on precise probability distributions [27] .

Recent studies highlight the challenges faced by commuters in major urban areas due to extended wait times and limited train capacity in crowded train stations [28] . The average waiting times for passengers vary based on factors such as station congestion, time of day, and trip purpose. The average waiting times for passengers vary based on factors such as station congestion, time of day, and trip purpose. During peak hours, passengers typically experience shorter wait times while random arrivals are more common during off-peak periods [20] [27] . Overcrowding can lead to longer waits at the original station as passengers vie for space on the train [29] . Furthermore, waiting time is a critical aspect of the passenger experience, significantly influencing behavior and potentially causing feelings of anxiety and dissatisfaction [30] . To address this issue, recent technological advancements have revolutionized the estimation of passenger waiting times. Smart card data has proven highly effective in collecting trip information at entry and exit points within the metro system, minimizing the need for manual data collection [31] .

Conversely, passengers who refer to schedules with varying arrival frequencies aim to minimize wait times by arriving close to departure times. Previous studies modeling passenger waiting times often assumed random arrivals and calculated average waiting times by multiplying the average bus headway by twice the ratio of the average headway to headway variance [32] . Initial investigations into estimating rail transportation wait times were influenced by research in bus transportation. Researchers explored models assuming uniformly distributed wait times using random passenger arrival models, advancing their studies in this area [33] [34] . A study conducted in Zurich, Switzerland, focused on measuring wait times for buses, trams, and trains at stations, recommending the use of a mixed uniform and Johnson SB distribution for modeling purposes [20] . Another study examining waiting times for both regional and metro lines in Copenhagen considered headway ranges from 2 to 60 minutes [35] .

Generally, the interval between train arrivals plays a crucial role in identifying potential conflicts during train headway and timetable optimization. Optimizing the metro system timetable is widely acknowledged as a traditional decision-making challenge that requires balancing the needs of passengers and the metro operating company [36] . Numerous research initiatives have aimed to enhance the functionality and productivity of public transportation systems. The metro timetable, closely linked to the train schedule, details arrival and departure times along with stop durations at each station. To enhance public transportation efficiency during peak hours, Beijing Metro Lines 1 and 2 have significantly reduced headway times to just two minutes. Consequently, it is imperative for operated trains to maintain consistent service frequency and adhere strictly to predetermine arrival and departure times specified in the schedule [37] . Moreover, fixed headway train timetables are inadequate for handling variable demand due to significant fluctuations in transportation needs [38] . To tackle this issue, a mixed-integer linear programming approach has been proposed [39] . Furthermore, a stochastic programming model has been proposed and refined for metro train rescheduling as a decision-making method to address railway routing and scheduling challenges [40] . Additionally, most train timetable optimization models consider the minimum headway between consecutive trains as a fixed value, although they also allow this minimum headway to be influenced by the current track assignment conditions [41] . Moreover, a mixed-integer linear programming model has been developed to optimize train schedules and reduce passenger waiting time disparities [42] .

The satisfaction and emotional responses of travelers are shaped by their perceptions and expectations of their journey, which are influenced by subjective feelings and views of various aspects of the travel experience [43] . Currently, there is no systematic method in place for evaluating passenger satisfaction at metro stations. Previous research has explored several factors contributing to passenger satisfaction, such as accessibility, information availability, time efficiency, customer service, comfort, safety, convenience, reliability, cost-effectiveness, and capacity considerations [44] . During the evaluation process, researchers typically choose indicators based on existing literature or personal experiences. However, these indicators can be subjective and may not accurately capture the evolving dynamics of metro services. Furthermore, opinions regarding the use of these indicators can vary depending on location, objectives, and time periods. Therefore, it is essential to carefully review and adjust evaluation criteria to align with specific research goals. Assessing passenger satisfaction with public transportation services is crucial for both transportation research and practical applications. To improve infrastructure, amenities, services, and increase public transport usage, transit agencies must understand how well they meet passenger expectations. Conducting customer surveys is critical as they provide valuable insights to transit agencies about aspects significant to passengers and specific areas of satisfaction or dissatisfaction [45] . In a study focused on Metro Rail Transit 3 (MRT3) stations in Metro Manila, Philippines, [46] identified reasons for low ridership from accessibility and inter-modality perspectives. The main sources of passenger dissatisfaction include station congestion, relatively high fares, and inconveniences in connecting transport facilities to other modes of transit. The remainder of this paper is organized as follows: Section 2 details the materials and method used. In Section 3, we present the main results of the study. Section 4 discusses these results and future research directions. Finally, Section 5 provide the conclusion.

The Apriori algorithm is a data mining technique used to discover patterns and associations within a dataset. It identifies frequent item sets and generates association rules based on these item sets. The algorithm evaluates the strength and relevance of the generated association rules using support, confidence, and lift measurements. The algorithm operates by incrementally enlarging the item sets until no additional frequent item sets can be discovered. It uses the following measures of significance and interest:

The specifics of the algorithm are as follows: Let $I=\left\{i_1,i_2,\cdots ,i_n\right\}$ represent the set of factors influencing metro passenger arrival volume referred to as the item set, and $D=\left\{t_1,t_2,\cdots ,t_n\right\}$ denote the set of responses from individual respondents, known as the data set; every response in D possesses a unique identifier and includes a subset of the items in I. A rule of an item set is expressed as {X} ⇒ {Y} where: $X,Y\subseteq I$ and $X\cap Y=\phi$ (X and Y are disjoint items). The sets of items X are called antecedent (or Left-hand side LHS) and sets of items Y consequent (or Right-hand side RHS) of the rules. There are mainly three measures of significance and interest which are the support, confidence, and lift. The support S(x) of an item set is the proportion of responses in the dataset which contains the item set given as:

$S\left(x\right)=\frac{f\left(x\right)}{N}$ (1)

$S\left(x\right)=\frac{f\left(y\right)}{N}$ (2)

$S\left(x\Rightarrow y\right)=\frac{f\left(x\cap y\right)}{N}$ (3)

where:

f(x) = Number of Instances with x

f(y) = Number of Instances with y

N = Total number of Instances

$f\left(x\cap y\right)$ = Number of Instances with both x and y

S(x) = Support of x item set

S(y) = Support of y item set

$S\left(x\Rightarrow y\right)$ = Support of the association

The confidence is an estimate of probability P(xy) of finding Consequent (RHS) of the rule in instances under the condition that these instances also contain the antecedent (LHS). The confidence is given as:

$\text{Confidence}\left(x\Rightarrow y\right)=\frac{S\left(x\Rightarrow y\right)}{S\left(x\right)}$ (4)

The Lift is the deviation of the support of the whole rule from the support expected under independence given the support of LHS and RHS, with a greater value indicating a better association.

$\text{Lift}\left(x\Rightarrow y\right)=\frac{S\left(x\Rightarrow y\right)}{S\left(x\right)S\left(y\right)}$ (5)

The Apriori algorithm, implemented using the open-source R programming language and the “arules” and “arulesviz” packages, was employed to investigate the relationship between high, moderate and low metro passenger arrival, with other factors such as waiting time, weather, and time of day, arrival pattern, travel behaviour and metro satisfaction.

Neural network models were deployed to detect nonlinear patterns in the data. The neural network architecture featured multiple hidden layers, and the training process incorporated backpropagation and the Adam optimizer. The effectiveness of the neural network models was assessed using metrics such as accuracy, precision, recall, and F1-score. The input data is transmitted to the hidden layers for processing, and the final hidden layer forwards the processed information to the output layer and receives the outcomes. This investigation employs a fully connected neural network, where each neuron in one layer is connected sequentially to every neuron in the subsequent layers, encompassing the input, hidden, and output layers. This approach is consistent with studies by [50] [51] . The process of deriving output data is delineated by the following equation:

$Y_k^{n+1}=f\left(\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{X}_i^n{w}_{ki}^n+b_i^n\right)$ (6)

where, $Y_k^{n+1}$X is output of unit k in the nth layer, f is the function of activation, $X_i^n$ is the input vector, $w_{ki}^n$ is a weight vector, $b_i^n$ is the bias weight.

The initial weights are typically assigned randomly at the beginning of neural network training, which involves adjusting these weights through backpropagation comprises two main phases: feedforward and backward propagation. A training set, consisting of input vectors and corresponding target output vectors, is provided to the network for learning. The network’s actual output is compared with the target output to calculate an error, which is then used to update the weights by propagating them. Iterative weight adjustments are performed for each training set until a stopping condition, such as a predefined number of epochs or a specified threshold, is met. The backpropagation algorithm consists of three key stages:

1) Feedforward Stage: The input layer computes the output by summing the weighted inputs and biases up to the output layer using a specified activation function.

2) Backpropagation stage: The error, obtained by comparing the network output with the target output, is calculated and propagated backward through the network starting from the output layer.

3) Weight and Bias Update Stage: In this final phase, the weights are adjusted to minimize errors based on the back propagated error signals.

Evaluation metrics were used to compare the performances of linear regression, neural network models, and association rule mining to determine the most effective approach for predicting passenger arrival. The complete methodology process is presented in Figure 1 below.

During a period of six weeks, we collected passenger arrival volume data at Yujiatou, Qingnian Road, and Jiangshe 2nd Road stations, focusing on both weekday and weekend passenger arrival volume during morning peak hours, off-peak hours, and evening peak hours. The data reveals some significant patterns. You can find the summary of this data in the provided Table 2 and Figure 2 , Figure 3 and Figure 4 below. For a comprehensive view of the entire dataset, which includes all stations and specific time intervals, please refer to the Appendix A.

The pie charts in Figure 2 show that weekday passenger volumes are consistently higher than weekends, reflecting typical commuter patterns. During summer holidays, more off-peak hour travel on weekends indicates increased leisure travel. This highlights the impact of summer holidays on metro travel patterns, with heightened off-peak and weekend travel volumes due to non-commuter activities.

Figure 3 Shows that In July 2023, there was increased off-peak hour and weekend travel, indicating more leisure travel due to school closures and vacations. In May 2024, there were dominant weekday morning peak hours and increased evening peak hour travel, suggesting more evening activities or later commutes.

Figure 4 above shows typical weekdays at the Qingnian Road Station. The higher percentage of evening peak hour dominance suggests that many commuters used this station to return home from work or school. In contrast, the lower off-peak percentage indicates that the station primarily serves as a hub for commuting, with less leisure or non-work-related travel occurring during the middle of the day. Compared to the Qingnian Road Station, the Jiangshe 2nd Road Station I had a noticeable increase in off-peak travel during weekends. A higher off-peak percentage on weekends indicates more leisure travel, with individuals using the station for activities such as shopping, dining, or visiting friends and family. However, evening peak hours remained dominant, suggesting significant evening activity.

Table 3 and Figure 5 , Figure 6 and Figure 7 below, shows a summarized information on significant fluctuations in average daily passenger arrival volume over several months, with peaks near 9000 passengers and troughs around 3000. Weekdays have a much higher average daily passenger arrival volume (approximately 5886 passengers) compared to weekends (around 1765 passengers), indicating the dominance of work and school commutes. Passengers flow is relatively balanced between mornings and evening on weekends, reflecting more flexible travel patterns. The detailed information is in the Appendix B.

Source: LAMATA Lagos Nigeria.

Figure 5 and Figure 6 show significant fluctuations in average daily passenger arrival volume over several months, with peaks nearing 9000 passengers and troughs around 3000. Weekdays have a much higher average daily passenger arrival volume (approximately 5886 passengers) than weekends (around 1765 passengers). Figure 7 reinforces this, showing a more significant portion of weekday passenger arrival volume. This emphasizes the need for optimized metro services to accommodate higher weekday demand while maintaining balanced weekend operations.

The Apriori algorithm analyzed metro passenger arrival volumes, generating 420 association rules grouped into high, moderate, and low classes. Strong correlations were found between high passenger volumes and longer waiting times, increased evening peak usage, high metro satisfaction, shorter waiting times, perceived waiting times, and weather conditions. Moderate volumes had significant associations with longer waiting times, disagreement on waiting times, neutral stance on perceived waiting times, and disagreement on weather conditions affecting arrival volumes. Low volumes strongly disagreed with perceived waiting times, shorter waiting times, overall metro satisfaction, increased evening peak usage, and weather conditions, all of which were strongly linked to low passenger volumes. These results highlight the impact of waiting times, peak periods, satisfaction levels, and weather conditions on metro passenger volumes. As presented in Table 4 and Figure 8 , Figure 9 and Figure 10 below.

Figure 8 above shows a strong correlation between passenger volumes and factors such as waiting times, weather impact, and peak hour usage at metro stations. It uses color intensity to represent the strength of these associations visually. Darker colors indicate stronger connections, while lighter colors indicate weaker ones. This visualization helps identify the factors most strongly correlated with high passenger volumes at metro stations.

From Figure 9 above, the visualization highlights that waiting times (both agreement and neutral stances) and weather conditions (neutral and disagreement stances) are significant factors that influence moderate passenger volumes at metro stations.

The figure highlights that the strongest correlations with low passenger volumes are between “WAT2 = Strongly Disagree” and “WAT3 = Strongly Disagree,” indicated by the highest lift values. Moderate correlations are seen in relationships such as “WTH1 = Strongly Disagree” and “MSS2 = Strongly Disagree.” Meanwhile, “WTH2 = Strongly Disagree” and “WTH3 = Extreme heat” show weaker, still significant, associations with lower lift values.

In the forthcoming analysis, Table 5 provides a comprehensive summary of the socio-demographic data obtained for this study. Three hundred sixty-five valid responses were collected from participants in Lagos, Nigeria. Among the respondents, 48.2% were male and 51.8% were female. The age distribution of the participants is as follows: under 18 years (19.2%), 18 years old (28.2%), 18 - 25 years (18.9%), 26 - 35 years (24.7%), and 36 - 45 years (9.0%). Regarding educational background, 46.8% of the participants are high school students, 36.4% are bachelor’s students, 11.5% are master’s students, and the remaining 5.2% are pursuing doctoral degrees. Additionally, 403 valid responses were obtained from participants in Wuhan, China. Among them, 48.4% were male and 51.6% were female. The age distribution in this group is as follows: under 18 years (18.9%), 18 years old (27.3%), 18 - 25 years (19.6%), 26 - 35 years (25.6%), and 36 - 45 years (8.7%). Regarding educational status, 47.6% of the participants are high school students, 30.0% are pursuing bachelor’s degrees, 11.4% are master’s students, and 5.0% are working towards doctoral degrees.

Neural networks are a type of machine learning model that closely resembles the human brain in structure and function. They excel at processing complex patterns and large amounts of data, making them particularly well-suited for tasks such as image recognition, natural language processing, and predictive analytics.

A lower MSE indicates superior model performance because it signifies that the predictions are closely aligned with the actual values. In this instance, an MSE of 0.1493 suggests that the model’s predictions are fairly accurate, although there is still potential for enhancement. R-squared values range from 0 to 1, with higher values denoting better model performance. An R-squared value of 0.8666 implies that approximately 86.66% of the variation in the dependent variable (Passenger Arrival) can be accounted for by the model. This suggests a robust association between the predictors and response variables. As presented in the Table 6 below.

In this instance, feature importance is calculated using a method that involves shuffling the values of each feature and then measuring the increase in the model’s error. The greater the error increase, the more significant the feature. The results indicate that waiting time was the most crucial feature in this model, with an importance score of 7.0581, signifying its substantial contribution to prediction accuracy, followed by MetroSatisfaction, weather conditions, time of day, ArrivalPattern, and general travel behavior. As presented in Table 7 and Figure 11 below.

The aforementioned Figure 11 illustrates the relationship between various features and passenger arrivals. Among these, Waiting Time is the most influential factor, implying that reducing waiting time can substantially enhance passenger arrival. Similarly, Metro Satisfaction has a high level of influence, suggesting that improving overall satisfaction can have a positive effect on passenger arrival. The impact of the weather was considered moderate, highlighting the importance of implementing weather adaptation strategies. The time of day also had a moderate effect, emphasizing the significance of optimizing operations based on time-of-day data. Finally, arrival patterns and travel behavior have a moderate influence, indicating the need for further analysis and a deeper understanding of these factors.

The model’s predictions were impressively accurate, with minimal residuals for each prediction. The model consistently demonstrated a balanced mix of positive and negative residuals, indicating well-calibrated performance. Additionally, there were no discernible patterns in the residuals, suggesting that the model effectively captured the underlying data patterns without overfitting or under fitting. As presented in Table 8 and Figure 12 and Figure 13 below.

The predicted value plot above shows the target variable’s forecasted values and the residual plot displays variances between actual and predicted values. The scatter plot indicates that the residuals are randomly distributed without any systematic pattern, confirming the model’s high accuracy and absence of significant biases.

Neural networks are a type of machine learning model that closely resembles the human brain in structure and function. They excel at processing complex patterns and large amounts of data, making them particularly well-suited for tasks such as image recognition, natural language processing, and predictive analytics.

The regression metric results provide an evaluation of the model’s performance. The Root Mean Squared Error (RMSE) is 0.1970, indicating the average magnitude of the prediction errors, with lower values suggesting better model performance. The Mean Absolute Error (MAE) is 0.0921, representing the average absolute difference between the predicted and actual values, with smaller values indicating more accurate predictions. The R-squared value is 0.5615, which means that approximately 56.15% of the variance in the target variable is explained by the model. This suggests a moderate level of explanatory power, indicating that while the model captures some of the variability in the data, there is still room for improvement. As presented in Table 9 below.

The importance of each feature is determined by its influence on the model’s predictions, known as feature importance. A higher value indicates a more significant impact on the model’s output. The feature “ArrivalPattern” has the highest importance score of 0.0659, “Weather” follows with a score of 0.039, “Time of day” has a score of 0.0331, “TravelBehaviour” with a score of 0.0322, and “MetroSatisfaction” has the lowest importance score of 0.0277. As presented in Table 10 and Figure 14 below.

The radar chart above visually represents the relative importance of five features (Weather, Time of day, arrival pattern, travel behavior, and MetroSatisfaction) in predicting the target variable. Each feature is plotted on an axis radiating from the center, with importance scores ranging from 0 to 0.07. The polygon formed by connecting the data points shows that ArrivalPattern has the highest importance, followed by Weather and Timeofday, while TravelBehaviour and MetroSatisfaction have lower importance scores. The chart’s shape and size provide a quick visual comparison, highlighting the most influential features clearly and intuitively.

The Predicted and Residual Results table compares the predicted and actual values for passenger arrival, demonstrating that the neural network model’s predictions are generally close to the exact values. For instance, the expected value of 2.8202 is near the actual value of 2.6667, and the predicted value of 3.0717 is close to 3.0000. The minor discrepancies between the predicted and actual values, known as residuals, indicate slight overestimations by the model, such as a residual of 0.1535 for the first row and 0.0717 for the second row. These small residuals suggest that the model’s predictions are pretty accurate, but there is still room for improvement. The consistency of the residuals across different rows indicates that the model is systematically close to the actual values, but fine-tuning the model or incorporating additional features could further enhance its predictive accuracy. As presented in Table 11 and Figure 15 below.

The scatter plot includes more prominent, semi-transparent blue points to represent the model’s predictions, a red dashed line indicating the ideal scenario where predicted values match actual values, and a green solid line showing the linear regression fit. The plot also features a title, subtitle, and improved axis labels for better readability.

However, the feature Arrival Pattern shows positive and negative impacts on the model’s predictions, with contributions ranging from −0.0565 to 0.7223 and a variable value of 0.03585, indicating its variability in influencing passenger arrival. Metro Satisfaction consistently has a negative impact, with contributions ranging from −0.7273 to −0.7105 and a variable value of −0.9683, suggesting that lower satisfaction decreases passenger arrival predictions. Time of Day consistently has a positive impact, with contributions ranging from 0.0057 to 0.4969 and a variable value of 1.081, indicating that certain times of the day increase passenger arrival predictions. Travel Behaviour consistently has a negative impact, with contributions ranging from −0.2914 to −1.4838 and a variable value of 1.306, suggesting that certain travel behaviors decrease passenger arrival predictions. Lastly, Weather consistently has a positive impact, with contributions ranging from 0.1149 to 0.2093 and a variable value of 0.6323, indicating that certain weather conditions increase passenger arrival predictions. As presented in Table 12 and Figure 16 below.

The bar plot above, shows the variable values for each feature, with colors indicating the sign of the contribution (positive, negative, or both). The contribution range is annotated above each bar, providing additional context on each feature’s impact variability. This visualization helps summarize and compare each feature’s influence on the model’s predictions. Please let me know if you have any further questions or need additional analysis.

High passenger volumes are commonly associated with extended waiting times, weather-related disruptions, and peak-hour usage. On the other hand, moderate passenger volumes tend to have neutral stances on waiting times and weather impact. In contrast, low passenger volumes are strongly correlated with strong disagreement on waiting times, weather impact, and metro satisfaction. The use of color intensity and point size in visualizing association rules offers a clear and concise understanding of the strength of the relationships between these various factors and passenger traffic volumes.

A Mean Squared Error (MSE) measure of 0.149 indicates that the model’s predictions were fairly precise. The R-squared value of 0.866 suggests that the model can account for approximately 86.6% of the variation in passenger arrival volume. Among the features, waiting time had the highest importance score of 7.058, followed by metro satisfaction, with a score of 6.119 and weather with a score of 1.031. Although, time of day, arrival pattern, and travel behavior also played a significant role, their importance scores were lower. The model predictions were remarkably close to the actual values, with residuals ranging between -0.007 and 0.035 units for the Lagos Blue Light Rail Line.

However, for the Wuhan data set, the R-squared value of 0.561 suggests that the model can account for approximately 56.1% of the variation in passenger arrival volume. Among the features Arrival pattern had the highest importance score of 0.065, weather had 0.039 followed by time of day, with a score of 0.033. The expected value of 2.820 is near the actual value of 2.666, and the predicted value of 3.071 is close to 3.000, suggesting that the neural network model’s predictions are generally close to the exact values.

After a comprehensive analysis of the two models, it is clear that the neural network model surpasses in terms of predictive capabilities due to its superior accuracy and reduced error rates. Additionally, association rule mining provides additional insights by discovering intricate relationships that might not be apparent through conventional regression analysis. The neural network model’s superior predictive performance, as demonstrated by its increased accuracy and decreased error rates, makes it the preferred choice among the evaluated models. With its more dependable and precise approach to forecasting passenger arrival volumes, this model can be invaluable for transportation planning and operations.

Table A1. Direct field observation of metro passenger arrival volume.

Date: 15^th-21^st May 2023 (Week 1)

Yujiatou Station

Date: 20^th -26^th May 2024 (Week 4)

Date: 27^th May-02 June 2024 (Week 5)

Qingnian Road Station

Date: 03rd - 09th June 2024 (Week 6)

Jiangshe 2nd Road

Table B1. Lagos light rail blue line passenger arrival volume statistics and revenue.

Table C1. Questionnaire content.

PAR = Passenger Arrival Volume, TOD = Time of Day, W = Weather, WAT = Waiting Time, MSS = Metro Satisfaction, TB = Travel Behaviour, AP = Arrival Pattern.